Multimedia Systems

https://doi.org/10.1007/s00530-020-00699-w

REGULAR PAPER

Video compression using hybrid hexagon search

and teaching–learning‑based optimization technique for 3D
reconstruction
B. Veerasamy1 · S. Annadurai2

Received: 4 April 2020 / Accepted: 22 September 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Motion estimation from a video sequence is an interesting issue in video processing. Nowadays, research has been focused
on global optimization techniques, that estimate the optical flow for pixel neighborhoods. In this paper, a hybrid statistically
effective motion estimation procedure has been proposed for better effectiveness video compression. This method explores
by utilizing a hexagonal search pattern with a secure number of search points at every lattice. It uses the association among
bordering pixels within the frame. So as to diminish the computative intricacy, this methodology uses hybrid hexagon search
and teaching–learning based optimization algorithm. This method additionally decreases the computational unpredictabil-
ity of block matching procedure. The image quality has been confirmed through 3D reconstruction using structured light
techniques. This strategy has been contrasted with different existing strategies and hereby utilizing the hexagon search-
based teaching–learning optimization algorithm could get a higher precision interms of PSNR of 44.36%, MSE of 2.40 and
compression ratio of 7.50.

Keywords Motion estimation · Hexagonal search · Compression · Block-matching algorithm · Peak signal to noise ratio

1 Introduction into squares [2]. In each square one search for the best coor-
dinating metric will be taken contrasted with the reference
Motion estimation is one of the most significant principal frame so the coordinating metric will be diminished. In this
issues in video compression. It has wide applications in the way, two frames of compression can be done by taking the
fields of robot route, blended and enlarged reality, visual best match. Among early block-based motion estimation,
tracking, image and video processing, astute transportation fixed search algorithms have been utilized. These calcula-
systems, etc. [1]. Motion estimation is a long way from a tions are used to lessen the computational prerequisite by
tackled issue, and it is as yet one of the dynamic research checking just a few points inside the search window while
topics in and beyond the computer vision community. Block- keeping a decent error execution when contrasted and Full
based motion estimation partition the input image frames Search algorithm. To diminish the search points without the
debasement of the image quality, block-based motion esti-
mation algorithms have been used. It mainly depends on the
Communicated by Y. Zhang.
motion vector conveyance of genuine video arrangements.
* B. Veerasamy It comprises two inquiry Patterns namely a large diamond
[email protected] search pattern and a small diamond search pattern. The high
S. Annadurai relationship between progressive frames of a video Sequence
[email protected] makes it conceivable to accomplish high coding effective-
1
ness by expelling the transient excess.
Department of Electronics and Communication Engineering, Numerous computationally proficient algorithms such
Hindusthan College of Engineering and Technology,
Coimbatore, Tamil Nadu, India as a three-step search, a new three-step search, a four-step
2 search, and diamond-based searching methodologies were
Department of Computer Science, Hindusthan College
of Engineering and Technology, Coimbatore, Tamil Nadu, developed for ME [2, 3]. These kinds of methodologies
India have been simulated and tested with different sizes. But the

13
Vol.:(0123456789)
 B. Veerasamy, S. Annadurai

diamond-based searching methodology embraces a diamond sub-squares. The principal sub-square was transformed by
search pattern, it has exhibited quicker preparing with com- discrete cosine transform, while the other three sub-squares
parative contortion with three-step search, new three-step were de-connected by transform lattices that were evalu-
search, and four-step searches [4–7]. In this paper, hexa- ated from the main remade sub-square by singular value
gon-based hybrid teaching–learning optimization has been decomposition. In [13], proposed to build up a preparation
implemented to accomplish considerable rapidity upgrad- dataset via looking for comparative squares in an area that
ing over the diamond search with comparative mutilation are utilized to compute transform lattices. Since looking
outcomes. through a comparative substance is computationally costly,
The remaining part of this paper is divided in the follow- it incredibly expands encoding and interpreting intricacy. In
ing way. A brief review of video compression techniques is [14], Pixel Motion CNN has been utilized which includes
presented in Sect. 2. Motion estimation and the approaches hybrid prediction networks. It demonstrates spatiotemporal
used for motion estimation, drawbacks of diamond search cognizance to successfully perform prescient coding inside
patterns and the hexagon patterns are discussed in Sect. 3. the learning system. Based on pixel motion CNN, authors
The overall workflow of this paper and the proposed method investigate a learning-based system for video compression
is clearly explained in Sect. 4. The simulation results are with extra segments of frequentative examination, blend, and
discussed in Sect. 5 and the conclusion of this paper is given so forth. Exploratory outcomes exhibit the adequacy of the
in Sect. 6. proposed plan. This learning-based plan gives a conceivable
new heading to additional improvement in the compression
ratio and functionalizes of video compression.
2 Literature review Learning-Based Image Compression techniques have
also been used [15–17] with kinds of neural network sys-
In [8], the complex motion could be depicted by smaller tems showing another bearing of video compression. For
squares while stationery substance or uniform motion by instance, the main exertion of learning-based image com-
bigger squares. However, the choice of a proper square is pression [18] was presented and shows their better execution
probably the most serious issue in applying workmanship contrasted with the primary encoding techniques. However,
standards to the environment and this offers to ascend to all learning-based strategies proposed so far were produced
an intriguing enhancement issue without a productive solu- for still image compression and there is still no distributed
tion for this issue. This issue has been overwhelmed by uti- work for video compression. The movement remuneration
lizing Variable Size Block Matching calculations. In [9], is also a successful device for video coding, can barely be
the authors exhibited ideal edge threshold-based quadtree prepared into a neural [19]. Various optimization techniques
movement estimator decomposition for movement estima- such as Genetic Algorithm, Differential Evolution, Harmony
tion by changing Edge Threshold to accomplish exchange off Search, Ant Colony Optimization, Shuffled Frog Leaping
between computational unpredictability and video quality in and Gravitational Search Algorithms have also been used
video compression. To take care of the issue of bit assign- for optimization problems in video compression. The cur-
ment, Lagrangian Multiplier updated calculations are uti- rent works in video compression mainly focus on improving
lized monumental different imperatives such rate and distor- compression ratio proficiency by independently heightening
tion for discovering multi limitation shortest paths. A* prune the quantization to diminish the pixel repetition. To address
methodology gives the best ideal solution for the R-D opti- this issue, in this paper, a hybrid teaching-learning-based
mization issue of video compression by choosing K-multi optimization algorithm has been used along with the HS
limitation shortest paths. In [10], a transform-based quan- pattern for the optimization so as to improve the efficiency
tization procedure has been introduced for Video compres- of the estimation purposes.
sion. Initially, content-dependent transform and Karhunen-
Loéve transform has been applied on the video sequence.
Then Jointly optimize transform and quantization has been 3 Motion estimation
performed for the enhancement issue and acquire a doable
solution for the improvement of coding execution. At last, a The motion estimation in video compression figures the
fast algorithm has been used to decrease the computational removal among the present edge and put away the past edge
unpredictability of jointly optimize transform and quantiza- that is utilized as the locus frame. Normally the prompt pre-
tion. In [11], the creators developed motion-compensated vious casing is viewed as the locus frame. Later video com-
estimation with some specific and after that move it on a pression gauges, for example, the H.264 bid adaptability in
level plane to create preparing database utilized to process choosing the references outlines and mixes could be picked.
transform networks. In [12], the authors developed to down- Figure 1 outlines the essential way of thinking of motion
sample a remaining square into four equivalent estimated estimation. Select a pixel having a place with the present

13
Video compression using hybrid hexagon search and teaching–learning-based optimization…

additionally presented as far as the flatness of motion vector

in the area. The flatness imperative creates the calculation
intuitive and entails too much huge calculation time, making
it unacceptable for pragmatic and constant execution. An
option and quicker methodology is the block-based motion
estimation. In this strategy, the applicant’s edge is isolated
into un-correlated squares of size for each frame, the better
motion vector is resolved in the loci frame. Here, a solitary
motion vector is figured for the whole block, where an innate
presumption that the whole square experiences translational
movement. This supposition that is sensibly substantial,
with the exception of the article limits and littler square size
prompts better motion estimation and compression.
Fig. 1  Motion estimation process
3.2 Annotations on diamond search pattern

edge, in relationship with its neighborhood as the up-and- It is realized that the search pattern has a significant impact
comers and after that decide its good coordinating location on rapidity and falsification performance in block-based
in the loci frame. The distinction in location among the up- motion estimation. Various sizes of search patterns are
and-comers and its match in the reference frame is character- regularly utilized in quick motion estimation techniques.
ized as the dislodging vector or all the more ordinarily, the Presently, diamond search patterns have been demonstrated
movement vector. In the wake of deciding the motion vector and be an increasingly effective alternative. These types of
one could foresee the present edge by smearing the removals patterns could discover huge movement squares with fewer
relating to the motion vector on the loci frame. focus points and furthermore diminish its vulnerability to
stalling out in nearby targets because of its generally huge
3.1 Motion estimation approaches advance extent in flat and perpendicular bearings. The
smaller state of the diamond search pattern about the center
Motion estimation can be broadly classified into two types, additionally produces less search patterns than four-step
they are (i) Pixel-based motion estimation and (ii) Block- search for discovering motionless or small motion vectors.
based motion estimation. The pixel-based motion estimation In a large diamond search in Fig. 1a, 8 search pattern has
procedure looks to decide the motion vector for each pixel various detachments from the middle argument has been
in the image. It is likewise alluded to as the visual stream utilized.
strategy, which deals with the essential supposition of bril- Figure 2b, c display the search procedure by utilizing
liance consistency, that is the power of a pixel stays steady the crook and edge points individually. In Fig. 2b, the point
when it is uprooted. However, no special counterpart for a noticeable “5” wins with littlest square alteration in the pre-
pixel in the reference edge is found toward the path ordinary sent stage and additionally, 5 more newly pursuit focuses
to the intensity gradient. It is hence that an extra limitation is (5–13) marked structure will be assessed in the following

Fig. 2  a Large diamond search pattern. b diamond search with crook point. c Diamond search with edge point

13
 B. Veerasamy, S. Annadurai

stage. In Fig. 2c, the point noticeable "4" triumphs in the not have the same number of frames, it will differ for each
present advance and 3 new up-and-comer focus "9," "10," different video sequences. The converted input video frames
and "11" will be checked in the accompanying advance. are applied as the input of hexagon-based hybrid teaching
From Fig. 2b, point "9" is closest to point "4" with separa- and learning based optimization methods. There, a hexa-
tion instead of the triumphant point "5" with separation 2. gon search pattern has been performed on the input video
On the off chance that the point "9" is the base or closest to sequences for motion estimation. Furtherly, so as to improve
the base among the search points appeared in the figure, the the exactness of the videos hybrid teaching and learning
twisting for the point "4" should doubtlessly be littler than based optimization method has been used as the optimiza-
that for point "5" as per the suspicion. Thusly, the point tion technique. This technique is utilized to improve the pre-
checked "9" in Fig. 2b is probably not going to be a champ cision of overall hexagon-based hybrid teaching and learn-
in the subsequent stage. A similar case applies to the point ing based optimization methods. The teaching and learning
checked "13". Similarly, the focuses stamped "9" and "11" based optimization strategy functioning based on the coach-
in Fig. 2c are additionally not a great possibility for the sub- ing- learning process. This technique depends on the impact
sequent stage search. Besides, the diamond search pattern of the encouragement of an instructor on the outcomes of
is so conservative as far as separation among surrounding students in a session. The outcomes of class are measured
focuses that there may exist some repetition between the as far as outcomes and evaluations. The educator must be
inquiry focuses, particularly at the start of lesser resolve a well-educated individual offers his or her insight with the
search. Subsequently, such dispersion of inquiry focuses students. The nature of an instructor’s ability influences the
on diamond search pattern is unproductive in discovering result of students. Clearly, a decent instructor trains the stu-
potential up-and-comers in the following stage. dent with the end goal that they can have better outcomes
The cause for the drawbacks previously mentioned for as far as their imprints or evaluations. Students additionally
diamond search is that the diamond search isn’t rough gain from the collaboration between themselves. This pro-
enough to a circle, which is simply 90° revolution of a cess will be continued till; there is no improvement at the
square. Preferably, a circle-molded pursuit design with a uni- outcomes of the hybrid teaching and learning based optimi-
form conveyance of a base number of hunt focuses is attrac- zation and it reaches the edge of the hunt window or 3 steps
tive to accomplish the quickest inquiry speed reliably. For afterward the harmonized condition.
all intents and purposes, a more circle-approximated search
design in the movement field is doable in which a base num- 4.1 Hexagon search pattern
ber of pursuit focuses are appropriated reliably. Each hunt
point can be correspondingly used with most extraordinary The fruitful presentation of the diamond search algorithm
viability, where the repetition among pursuit guides should uncovers the significance of the outline and scope of inquiry
be cleared maximally. design in motion estimation, in terms of conjunction rapidity
and falsification performance. The hexagon search algorithm
has also been developed for motion estimation in video pro-
4 Workflow of the proposed method cessing. The hexagon search has two examples. The bigger
one comprises of 7 search points with single focus encom-
The overall workflow of this paper is given in Fig. 3. Both passed by 6 points. The SHSP resembles a precious stone
the hexagon search pattern and hybrid teaching and learn- having four pursuits focuses on the inside with separation 1.
ing based optimization techniques have been utilized in the In the initial step, seven search points are utilized to
proposed method. Initially, the considered video sequences find the best-coordinated square. In the event that the ideal
are converted into the number of frames. Every video does is at the inside, the large hexagon search pattern will be

Fig. 3  The workflow of the

proposed method

13
Video compression using hybrid hexagon search and teaching–learning-based optimization…

supplanted by a small hexagon search pattern. Then four

( )
Nhex Px , Qy = 7 + 3 × k + 4,
checking points secured by this small hexagon search pattern
will be analyzed to decide the final solution with minimum where is Px, Qy is the final motion vector found, and ‘k’ the
block distortion. Something else, the inquiry system pro- quantity of enormous hexagon pattern.
ceeds by moving the focal point of the enormous hexagon
to the point with least square mutilation and assess the new
applicant focuses. Note that while the huge hexagon search 4.2 Proposed HS‑HTLBO algorithm
moves along the specified direction with diminishing square
bending, just three new non-overlapped search focuses will The flow diagram of hexagon search(HS)-hybrid teaching
be assessed each time. The search procedure of the hexagon and learning based optimization(HTLBO) has presented in
search algorithm may be depicted as follows. Fig. 5. The computationally productive motion estimation
technique utilizing the fixed number of SP at every system
Step 1. The hexagon with seven checking points are framework along with the hybrid teaching and learning
focused at (0, 0), the focal point of a predefined search based optimization strategies has been proposed for video
window. When the event that the best match point hap- compression. The remarkable highlights of hexagon search-
pened at the focus of the enormous hexagon, continue teaching and learning-based optimization are as per the fol-
to Step 3, otherwise to Step 2. lowing:—The diamond search pattern utilized in video com-
Stage 2. With the least mean square error point in the past pression uses a variable number of search points. When the
search step as the center, another huge hexagon is size of the sequences increased, the quantity of search points
framed with the three new competitor focuses and also upsurges. In the hexagon search-teaching and learning-
the nonoverlapped three old competitor focuses are based optimization algorithm, a fixed number of SP (6) are
checked. Proceed step 3, if the minimum mean square utilized. The utilization of the fixed number of focus points
error point remains the center point of the newly gen- at every matrix brings about the low multifaceted nature of
erated hexagon. Otherwise, continue step 2 until the hardware implementation in FPGA architecture at the equip-
required point is reached. ment level.
Stage 3. Switch the search pattern from the enormous to the The implementation of hexagon search-teaching and
small size of the hexagon. The four points secured by learning-based optimization (HS-HTLBO) is explained as
the small hexagon are assessed to contrast with the follows,
present least mean square error point. The new least
mean square error point is the last arrangement of the Step 1. Prepare the initial parameters Choose the no: of
movement vector. search points and optimization parameters. The opti-
mization problem is defined as,
Figure 4 demonstrates a case of the pursuit way meth- minimizeO(X)
odology prompting the movement vector, where 17 SP are
assessed in four stages successively. From the strategy, it
prompts a statement of the complete number of search points Here, O(X)required impartial function.
for each present square as,
The above process applies to each block in the current Step 2. HS It utilizes HS pattern with 7 SP for all separa-
frame for block motion estimation. From the procedure, it tions > 1. Based on the separation of the better-coor-
can be easily derived that the total number of search points dinated competitor, this algorithm will moreover be
per block will be: ended or complete the subsequent HTLBO optimiza-
tion for better precision.
Step 3. 2-point search This progression is executed just
when the separation of the best-coordinated square in
Solutions are improving the past advance is equivalent to one. It accomplishes
Output Frames once and gives the last most ideal coordinated up-and-
comer in the hunt window.
Input HS HTLBO Reconstruction Step 4. Teacher phase Estimate the average of each frame
Frames
No improvement A = [a1, a2, a3….. an]. The better-matched solution is
treated as a teacher for next reiteration OTeacher. Now
Compression the educator makes alteration from ‘A’ towards OTeacher
and it will be treated as the new average for the next
Fig. 4  Hexagon search pattern reiteration. Therefore, Anew = Oteacher. The changes in

13
 B. Veerasamy, S. Annadurai

Initialize number of search points and termination criterion

Perform hexagon search pattern

Yes best possible matched

block at distance=0?

No

Yes best possible matched

block at distance=1?

Search nearby two points

No
Calculate the mean of each frame

Identify the best solution

Calculate the difference mean and modify the solutions based on best solution

new solution Yes Accept

better than

No
Select the solutions randomly and modify them by comparing with each other

Is new solution Yes

Accept
better than

No
Replace worst solution with better solution

Satisfied No
termination

Yes
Final Solution

Fig. 5  Flow diagram of Hexagon search- teaching and learning based optimization algorithm

13
Video compression using hybrid hexagon search and teaching–learning-based optimization…

the average mean can be expressed as D = r (Anew-Tf × learning-based optimization techniques. The 3D recon-
A). Here Tf has been chosen within the range of (1–2). structed images are given in Fig. 8. The output images cap-
This value is further added with the existing resolution tured using structured light techniques must be rotated to get
to appraise its values utilizing Onew = Oold + D. Agree a clear view of a video. Here, the 3D reconstruction of the
to take Onew, if it provides a better solution than the single frame of every video has been presented.
previous solution. The 3D surface has some artifacts. This sort of artifacts
Step 5. Student phase The worst solution is replaced with not present in the original and decompressed 2D image at a
a better solution as explained above. So that the better lower compression ratio. Artifacts show up when the struc-
solution can be achieved for the next frame in the video tured light patterns are not unmistakably characterized in
sequence. the image, or debased after compression/decompression.
Step 6. Termination criterion Stop search either grasps the It is demonstrated that the structure lights are unmistak-
frontier of the search window or 3 steps subsequently ably noticeable at a higher compression ratio of 99%. In
better-matched condition. this paper, for the analytic purpose, the video sequences are
compressed and decompressed and the outcome frames are
changed over to a 3D surface. JPEG algorithm has likewise
5 Simulation results and discussion been utilized to compress/decompress the video succes-
sions. While the JPEG algorithm essentially neglected to
This method has been simulated for 14 input video sequences compress the video sequences at the required compression
which include both 2D and 3D video sequences. It is simu- ratio. Additionally, the significant point to note is that 2D
lated on MATLAB 2017a system configuration, i5 proces- and 3D recordings could not decompress to proportional
sors with 4 GB RAM. The metadata of video sequences are quality for 3D reconstruction.
given in Table 1 and the input video sequences are given in
Fig. 6. Initially, the input video sequences have been con- 5.1 Performance evaluation
verted into the number of input frames and these frames
have been used for compression. The assessment of performance metrics is a productive
The sequences of testing can be beneficial to compare method to evaluate the recital of hexagon search-teaching
the relative spatial info and temporal info found in the sev- and learning-based optimization with other existing tech-
eral sequences available. The compression factor is directly niques. Before one becomes acquainted with the recital of
proportional to the spatial and temporal information of a hexagon search-teaching and learning-based optimization
sequences. techniques, knowing extensively the meanings of metrics
The compressed images using the hexagon search- is inexorable. The performance of hexagon search-teach-
teaching and learning-based optimization method and other ing learning-based optimization strategy is investigated
methods are given in Fig. 7. Then the compressed images regarding some presentation estimates which include peak
are given to the 3D reconstruction. Here, the compressed signal to noise ratio, compression ratio and mean square
images are 3D reconstructed using a structured light tech- error with some current block matching algorithms. Peak
nique and compared with the hexagon search-teaching and signal to noise ratio is utilized to quantify the nature of the

Table 1  Metadata of video Sequence Format Total Frames Motion Type Annotations SI TI
sequences used for Simulations
Butterfly mp4(1920 × 1080) 339 Medium Video_a 154.5 35.1
Car mp4(596 × 336) 394 Medium Video_b 81.7 30.8
Frog mp4(596 × 336) 333 Medium Video_c 227.0 46.4
Parrot mp4(1920 × 1080) 515 Medium Video_d 160.3 34.7
Red Leaf mp4(596 × 336) 424 Low Video_e 157.8 36.1
Waterfall mp4(600 × 316) 345 High Video_f 79.7 26.8
Wolf mp4(596 × 336) 537 High Video_g 214.9 46.3
Leopard mp4(596 × 336) 477 Low Video_h 159.5 32.4
Fish mp4(596 × 336) 620 Medium Video_i 219.1 47.7
Coat mp4(640 × 360) 750 Medium Video_j 254.6 49.2
xCap mp4(600 × 316) 172 High Video_k 56.2 5.3
Akiyo mp4(352 × 264) 301 Medium Video_l 58.7 24.6
Flower mp4(596 × 336) 527 Medium Video_m 213.5 2.7

13
 B. Veerasamy, S. Annadurai

Fig. 6  Input video sequences

reconstructed compressed image. The higher value of peak Another parameter is the compression ratio. It is the ratio
signal to noise ratio (PSNR) indicates a better quality of the among the total number of bits in the compressed image
recreated image. and in the original image. Higher values of the compression
m−1 n−1
ratio indicate better compression of the original image. It is
1 ∑∑[ ]2 defined as,
MSE = I(i, j) − K(i, j) (1)
mn i=0 j=0
Number of bits in Compressed Image
Compression Ratio = .
Number of bits in original image
(4)
( )
MAX2I
PSNR = 10 log10 (2) The peak signal to noise ratio values of hexagon search-
MSE
teaching and learning-based optimization techniques with
� � other existing methods are listed in Table 2. From the table,
MAXI it can be noticed that hexagon search-teaching and learning-
= 20 log10 . (3)
based optimization methodology gives better peak signal
√
MSE
to noise ratio values when compared with other methods.
The diamond search produces very poor peak signal to noise

13
Video compression using hybrid hexagon search and teaching–learning-based optimization…

Fig. 7  Simulation outcomes of

HS-HTLBO and other existing
methods

13
 B. Veerasamy, S. Annadurai

Fig. 8  3D reconstructed images Performance scrutiny of HS-HTLBO

ratio values so that the quality of the compressed image will 3. PSNR and MSE values are highly correlated. The higher
be very poor using that method. value of mean square error yields poor peak signal to noise
The mean square error values of hexagon search-hybrid ratio value so that the quality of the compressed image will
teaching and learning based optimization, hexagon search, be poor. The proposed hexagon search-hybrid teaching and
diamond search-hybrid teaching and learning based opti- learning based optimization (HS-HTLBO) method give
mization and diamond search methods are given in Table

13
Video compression using hybrid hexagon search and teaching–learning-based optimization…

Table 2  Comparison of PSNR values of HS-HTLBO with other Table 4  Comparison of compression ratio values of HS-HTLBO with
existing methods other existing methods
Video Sequences HS-HTLBO HS DS-HTLBO DS Video sequences HS-HTLBO HS DS-HTLBO DS

Video_a 44.11 43.14 42.33 40.22 Video_a 7.39 6.12 5.96 4.16
Video_b 44.11 43.42 42.38 40.18 Video_b 7.79 6.48 5.53 5.45
Video_c 44.52 43.27 42.20 40.62 Video_c 7.67 6.27 5.68 4.32
Video_d 44.78 43.77 42.19 40.76 Video_d 7.50 6.36 4.23 4.98
Video_e 44.31 43.18 42.65 40.14 Video_e 7.69 6.23 5.61 4.52
Video_f 44.09 43.45 42.32 40.25 Video_f 7.42 6.52 5.98 4.65
Video_g 44.20 43.81 42.62 40.86 Video_g 7.97 6.24 5.99 4.87
Video_h 44.24 43.60 42.28 39.91 Video_h 7.48 6.01 5.41 4.82
Video_i 44.33 43.52 42.41 39.94 Video_i 7.19 6.41 4.25 4.11
Video_j 44.09 43.22 42.79 40.26 Video_j 7.81 6.32 6.24 5.23
Video_k 44.27 43.42 42.86 40.42 Video_k 7.35 6.08 4.89 4.78
Video_l 44.09 43.03 42.55 40.27 Video_l 7.30 6.19 5.75 5.01
Video_m 45.55 43.05 42.77 40.38 Video_m 7.14 6.26 5.66 4.56
Average 44.36 43.37 42.48 40.32 Average 7.51 6.26 5.47 4.72

Table 3  Comparison of MSE values of HS-HTLBO with other exist- The analysis of peak signal to noise ratio measure for the
ing methods thirteen video sequences is depicted in Fig. 9. The results
Video Sequences HS-HTLBO HS DS-HTLBO DS 3D conclude that the maximum peak signal to noise ratio value
is attained in the proposed hexagon search-hybrid teaching
Video_a 2.54 3.18 3.83 6.18 0.63 and learning-based optimization algorithm. Also, the values
Video_b 2.54 2.98 3.78 6.23 2.10 are compared with some existing block matching algorithms
Video_c 2.31 3.08 3.94 5.64 1.32
like hexagon search, diamond search-hybrid teaching learn-
Video_d 2.18 2.75 3.95 5.45 1.07
ing based optimization and diamond search. Figure 10 shows
Video_e 2.43 3.15 3.56 6.43 1.37
the error rate (MSE) analysis of proposed algorithm hexagon
Video_f 2.55 2.96 3.84 6.13 1.40
search-hybrid teaching and learning based optimization and
Video_g 2.49 2.72 3.58 5.33 0.89
existing techniques such as hexagon search, diamond search-
Video_h 2.46 2.86 3.87 6.63 1.07
hybrid teaching learning based optimization and diamond
Video_i 2.42 2.91 3.76 6.59 0.85
search. From Fig. 10, it can be easily observed that the hexa-
Video_j 2.55 3.12 3.44 6.12 1.99
gon search-hybrid teaching and learning based optimization
Video_k 2.45 2.98 3.39 5.89 1.15
method gives the lowest value of mean square error and the
Video_l 2.55 3.26 3.64 6.11 1.13
new three-step search method produces the highest values
Video_m 1.82 3.24 3.46 5.95 0.91
of mean square error.
Average 2.40 3.01 3.69 6.05 1.22
The graphical representation of comparative analy-
sis of another important parameter, the compression ratio
is depicted in Fig. 11. From the graph, it can be visually
the least mean square error when compared with other noticed that hexagon search-hybrid teaching and learning
existing methods. based optimization method gives a higher compression
The Compression ratio of hexagon search-hybrid teaching ratio when compared with all other existing methods which
and learning based optimization, hexagon search, diamond include hexagon search, diamond search-hybrid teaching,
search-hybrid teaching and learning based optimization and learning-based optimization, and diamond search.
diamond search methods are tabled in Table 4. The lower The average values of various performance evaluation
value of the compression ratio yields a poor compressed parameters of hexagon based- hybrid teaching and learning
image. The compression ratio must be high for better com- based optimization and other existing methods are calculated
pression. Here, hexagon search-hybrid teaching and learning and listed in Table 5. Hexagon based- hybrid teaching and
based optimization method produce the highest compression learning based optimization method produced the average
ratio when compared with another hexagon search, diamond peak signal to noise ratio of 44.36, the mean square error
search-hybrid teaching and learning based optimization and of 2.40 and the compression ratio of 7.51. After analyzing
diamond search methods. the values, the mean square error value is much low in of

13
 B. Veerasamy, S. Annadurai

Fig. 9  Peak signal to noise ratio

values PSNR Values
46
45
44
43
42
41
40
39
0 2 4 6 8 10 12 14

HS-HTLBO HS DS-HTLBO DS

Fig. 10  Mean square error

values MSE Values
7

0
0 2 4 6 8 10 12 14

HS-HTLBO HS DS-HTLBO DS

Fig. 11  Compression ratio

Values CR values
9

0
0 2 4 6 8 10 12 14

HS-HTLBO HS DS-HTLBO DS

hexagon based- hybrid teaching and learning based optimi- signal to noise ratio has been increased with 0.99% when
zation method and the peak signal to noise ratio, compres- compared with HS and 1.88% with diamond based- hybrid
sion ratio values have been increased in the proposed method teaching and learning based optimization methods. The
when compared with other existing methods. The peak graphical illustration of comparative analysis of hexagon

13
Video compression using hybrid hexagon search and teaching–learning-based optimization…

Table 5  Average values of performance parameters for hexagon Table 6  Average value of 2D Methods 2D MSE 3D MSE
search-hybrid teaching and learning based optimization and other and 3D Mean square error
existing methods (MSE) Average 3.78 1.22
Parameters → PSNR MSE CR
Methods
↓
Table 7  Analysis of perceptual evaluation of video quality (PEVQ)
HS-HTLBO 44.36 2.40 7.51 metrics
HS 43.37 3.01 6.26 Video sequences HS-HTLBO HS DS-HTLBO DS
DS-HTLBO 42.48 3.69 5.47
Video_a 4.43 3.65 3.42 3.12
DS 40.32 6.05 4.72
Video_b 4.44 3.45 3.32 3.23
Video_c 4.12 3.35 3.43 3.15
Video_d 4.53 3.65 3.56 3.26
based- hybrid teaching and learning based optimization and
Video_e 4.32 3.45 3.45 3.14
other existing methods have been presented in Fig. 12.
Video_f 4.23 3.55 3.55 3.16
The average value of 2D mean square error and 3D mean
Video_g 4.43 3.55 3.78 3.12
square error values are given in Table 6. The average value
Video_h 4.76 3.65 3.65 3.16
of 2D mean square error has been calculated by taking the
Video_i 4.65 3.65 3.56 3.12
average value of four compression methods which includes
Video_j 4.45 3.65 3.43 3.16
hexagon-based hybrid teaching learning optimization
Video_k 4.54 3.65 3.57 3.43
method, diamond-based hybrid teaching learning optimiza-
Video_l 4.64 3.65 3.65 3.12
tion method, hexagon search and diamond search methods.
Video_m 4.46 3.65 3.56 3.21
2D methods produce the average value of mean square error
Average 4.46 3.58 3.53 3.18
is 3.78 and 3D reconstruction produce the average mean
square error of 1.22. From the table, 3D reconstruction using
a structured light technique produces the smallest mean
square error. But the major drawback in this 3d reconstruc- table. The test was found to be significant and the HS-
tion is, it could not decompress to proportional quality for HTLBO yields efficient results, when compared with the
3D reconstruction or in the event that it does, the 3D surface other algorithms. The statistical analysis is represnetd in
contains some corruption. The ITU-T recommended per- Table 8.
ceptual evaluation of video quality (PEVQ) metrics for the The speed improvement in percentage of the proposed
evaluation of compression performance and the values of HS-HTLBO over HS, DS-HTLBO and DS is listed in
PEVQ metrics are represented in Table 7 below. The results Table 9. The speedup percentage for the video seuences
in table reveals that, HS-HTLBO is superior, when com- shows that the proposed method outperforms the existing
pared with other algorithms. methods.
The statistical analysis was performed by the
Mann–Whitney U test and the results are depicted in

Fig. 12  Comparative analysis of 50

HS-HTLBO and other existing
45
methods
40
35
30
25
20
15
10
5
0
PSNR MSE CR

HS-HTLBO HS DS-HTLBO DS

13
 B. Veerasamy, S. Annadurai

Table 8  Analysis of Mann–Whitney U test metrics contortion. A hexagon-based hybrid teaching and learning
Algorithms taken for analysis Performance analysis based optimization method strategy is additionally giving
with metrics under considera- great outcomes at a worthy corruption in picture quality.
U value Z Score Test result
tion The peak signal to noise ratio value has been increased
with 0.99% when compared with hexagon search and
HS-HTLBO and HS 43 4.4799 Significant
1.88% with diamond-based hybrid teaching and learning
HS-HTLBO and DS-HTLBO 47 4.4799 Significant
based optimization method methods. This paper demon-
HS-HTLBO and DS 47 4.4799 Significant
strates the performance analysis of proposed hexagon-
HS-HTLBO and HS 39 − 4.30769 Significant
based hybrid teaching and learning based optimization
HS-HTLBO and DS-HTLBO 39 − 4.30769 Significant
block matching algorithm for H.264 codec for dependable
HS-HTLBO and DS 39 − 4.30769 Significant
video compression as for bit-rate and gains in peak signal
HS-HTLBO and HS 38 − 4.5676 Significant
to noise ratio with less computation complexity.
HS-HTLBO and DS-HTLBO 38 − 4.5676 Significant
HS-HTLBO and DS 38 − 4.5676 Significant

Compliance with ethical standards

Table 9  Tiing Coplexity Analysis of HS-HTLBO over existing meth- Conflict of interest This paper has no conflict of interest.
ods
Video sequences HS-HTLBO HS-HTLBO over HS-
over HS DS-HTLBO HTLBO
over DS References
Video_a 45.7 25.7 58.1
1. Aziz, T., Dolly, D.R.J.: Motion estimation and motion compen-
Video_b 39.13 23.12 53.7 sated video compression using DCT and DWT. Int. J. Emergl
Video_c 41.13 24.5 56.01 Technol. Adv. Eng. 2(12), 667–671 (2014)
Video_d 39.13 25.1 51.7 2. Shiju, P.P., Chakrabarti, I., Virdi, R., Wasnik, S.: VLSI architec-
ture for fixed mesh based deformable motion estimation using
Video_e 42.43 24.6 58.9
ARPS algorithm. Microprocess. Microsyst. 59, 92–102 (2018)
Video_f 41.8 23.8 52.4 3. Díaz-Cortés M.A., Cuevas, E., Rojas, R.: Motion estimation
Video_g 47.1 26.16 54.9 algorithm using block-matching and harmony search optimi-
zation. In: Engineering Applications of Soft Computing, pp.
13–44. (2017)
4. Yu, W., Hu, D., Tian, N., Zhou, Z.: A novel search method based
6 Conclusion on artificial bee colony algorithm for block motion estimation.
EURASIP J. Image aVideo Process. Article number 2017, 66
This paper has presented and demonstrated a new method (2017)
for image compression and illustrated the quality of com- 5. Jianhua, Lu, Liou, M.L.: A simple and efficient search algorithm
for block-matching motion estimation. IEEE Trans. Circ. Syst.
pression through 2D and 3D reconstruction, 2D and 3D Video Technol. 7(2), 429–433 (1997)
root mean square error. A new quickest algorithm utilizing 6. Po, L.-M., Ma, W.-C.: A Novel Four-Step Search Algorithm for
a Hexagon search design in block-based motion estima- Fast Block Motion Estimation. IEEE Trans. Circ. Syst. Video
tion shows a substantial increase with the various types of Technol. 6(3), 313–317 (1996)
7. Li, R., Zeng, B., Liou, M.L.: New three-step search algorithm for
diamond search-based search algorithms while keeping block motion estimation. IEEE Trans. Circ. Syst. Video Technol.
up comparable contortion execution. Hybrid teaching and 4(4), 438–442 (1994)
learning based optimization algorithms is moderately a 8. Kumar, R., Kumar, K., Pandit, A.K.: Rate-distortion optimized
new algorithm and can possibly take care of the streamlin- motion estimation for video compression using improved quad
tree algorithm. Ekoloji 28(107), 17–25 (2019)
ing issues in motion estimation strategies. The proposed 9. Kumar, R., Kumar, K., Pandit, A.K.: Performance and complexity
hexagon-based hybrid teaching and learning based opti- analysis of motion estimation using multiple constraints in video
mization method could discover motion vector in move- compression. In: International Conference on Intelligent Comput-
ment with fewer search points over other diamond search ing and Communication Technologies, pp. 698–706, 2019.
10. Wang, M., Xie, W., Xiong, J., Wang, D., Qin, J.: Joint optimiza-
methods. The simulation outcomes also confirmed the tion of transform and quantization for high efficiency video cod-
proclamation, it has persuasively exhibited the prevalence ing. IEEE Access 7, 62534–62544 (2019)
of hexagon-based hybrid teaching and learning based opti- 11. Biswas, M., Pickering, M.R., Frater, M.R.: Improved H.264-based
mization method over the other strategies as far as uti- video coding using an adaptive transform. In: Proc. IEEE Int.
Conf. Image Process. (ICIP), Sep. 2010, pp. 165–168.
lizing the most modest number of pursuits focuses with 12. Wang, M., Ngan, K.N., Xu, L.: Efficient H.264/A video compres-
an extremely little punishment of minimal debasement in sion video coding with adaptive transforms. IEEE Trans. Mul-
timed. 16(4), 933–946 (2014)

13
Video compression using hybrid hexagon search and teaching–learning-based optimization…

13. Lan, C., Xu, J., Zeng, W., Shi, G., Wu, F.: ‘Variable block-sized 18. G. Toderici, S. M. O’Malley, S. J. Hwang, D. Vincent, D. Minnen,
signal dependent transform for video coding’. IEEE Trans. Circ. S. Baluja, M. Covell, and R. Sukthankar, “Variable rate image
Syst. Video Technol. 28(8), 1920–1933 (2017) compression with recurrent neural networks,” in International
14. Chen, Z., He, T., Jin, X., Wu, F.: Learning for video compression. Conference on Learning Representations (ICLR), 2016.
In: IEEE Transactions on Circuits and Systems for Video Technol- 19. Ohm, J., Wien, M.: Future video coding coding tools and develop-
ogy, pp. 1–12, 2019 ments beyond hevideo compression. In: Wien, M. (ed.) Tutorial
15. Toderici, G., Vincent, D., Johnston, N., Jin Hwang, S., Minnen, in International Conference on Image Processing (ICIP) (2017)
D., Shor, J., Covell, M.: Full resolution image compression with
recurrent neural networks. In: Computer Vision and Pattern Rec- Publisher’s Note Springer Nature remains neutral with regard to
ognition (CVPR), July 2017. jurisdictional claims in published maps and institutional affiliations.
16. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized
image ´ compression. In: International Conference on Learning
Representations (ICLR), 2017
17. Theis, L., Shi, W., Cunningham, A., Huszar, F.: Lossy image
com- ´ pression with compressive autoencoders. In: International
Conference on Learning Representations (ICLR), 2017

13