Image Classification Using Deep

Learning

Mohd. Suhail , Shivam , Vaibhav Bansal

Student of

M.C.A.GalgotiasUniversity

Dr. Shambu Kumar Jha

Assistant Professor, Galgotias University

Plot No.-2, Sector-17A, Yamuna Expressway, Gautam

Buddha Nagar,

Greater Noida, Uttar Pradesh, India

 1 Abstract: Image classification is just a way that 25 Introduction:- As reported by the Internet Data
2 will balance the visual dissimilarity between 26 Center (IDC), the entirety of global data
3 machines and humans, which will help the 27 It will add up to 45ZB by 2020. Much of it, some
4 computer to interpret an image in a way as 28 70%, will seem like images and videos. From
5 humans perceive. 29 computer vision, big data can now be extracted
6 Despite its rampant application in practice, the 30 from these pictures and videos. Computer vision
7 conventional methodology has been plagued by 31 is getting more and more common these days in
8 numerous problems. Its accurate classification 32 the picture classification space. Since its
9 precision is highly and artificially limited while 33 development in the late 1950s, image
10 unsatisfactory results are yielded. Thus, deep 34 classification has been widely used in numerous
11 learning has been put forward as a possible 35 engineering fields, the medical profession,
12 improvement to the accuracy of the solution to 36 human-car monitoring, fingerprint technology,
13 challenges involving image classification. 37 disaster management, security, and so on.
14 This article addresses fundamental principles 38
15 related to picture classification employing deep 39 Classification is the systematic structuring and
16 learning, examines different methodologies for 40 grouping of items based on its attributes. In an
17 image classification, highlights the benefits 41 attempt to bridge the perceptual chasm between
18 associated with deep learning, and explores the 42 man and machine, image categorization is used
19 diverse applications that utilize deep learning in 43 to provide data that enables machine learning.
20 the context of image classification.
21 Artificial neural networks, computational
22 learning, deep learning techniques, and visual
23 categorization.
24
44 images is based can be generalized into four 93 remove undesired data, carry out feature
45 categories. 94 extraction and filtering, and eventually image
46 95 classification by a machine learning
47 1. Image classification using statistical methods: 96 discrimination. As the applications of machine
48 This well-known statistical method is based on 97 vision in a variety of sectors continue to expand,
49 the error-minimizing principle and includes 98 people can now quickly perform tasks like image
50 Bayesian and Markov models. 99 processing methods such as noise reduction and
51 100 feature extraction.
52 2. Image classification based on textures,101
53 geographical traits, and traditional color schemes.102
54 103 Owing to the data-driven nature of the analysim
55 3. Image classification using shallow learning. 104 of ever-growing volumes of varying data,
56 105 transfer, and semi-supervised learning, as well as
57 4. Deep learning powered categorization by106 the automated construction and assessment of
58 images.Machine learning is the standard107 deep learning models using AI-driven methods,
59 technique for classifying images. A feature108 it is sometimes difficult to overemphasize the
60 extraction module in machine learning extracts109 role and significance of artificial intelligence in
61 significant features such as edges and110 this regard. It is no secret that AI-powered
62 texturessuggestion for categorizing 111 technologies are establishing themselves in
112 numerous areas of human activity, and most
63 5. In this work image classification using deep113 researchers or entrepreneurs claim that Every
64 learning is covered. For this task, the114 technology sector has its own ‘big expectation’,
65 convolutional neural network architecture115 so the artificial intelligence’s ‘big hope’ rests
66 AlexNet is used. For the purpose of116 many perception and notion. Artificial
67 classification, four images are taken from the117 Intelligence has experienced its fair share of ups
68 ImageNet database. Different experiments were118 and downs, yet it is still considered the most
69 conducted where the images had their portions119 promising technology of our time. It is still
70 cut out of different areas. The results demonstrate120 intricate how AI-powered software applications
71 the applicability of deep learning’s image121 stand in various industries and whether any AI
72 classification capabilities in AlexNet and show122 system would suit every business model, so AI
73 its reliability. 123 adaptation isn’t that widespread, however,
74 124 perception in regards to AI is ever developing
75 The first task in image recognition is capturing125 ever since ai came into those ears. coversa vast
76 the picture using some image obtaining devices,126 amount of previously unseen technologies which
77 and thereafter the picture data is processed to and127 piqued interest around AI.
78 a module that does the classification using the
79 features extracted. However, there are significant128 It then processes the pixel values of each image
80 restrictions regarding how machine learning may129 into an internal representation that would help the
81 be used for the purpose of photo classification.130 classifier recognize patterns in the input
82 The technique can only extract a specific set of131 photographs. There are several deep learning
83 features from the photos and cannot use features132 networks, including the following:
84 that are not those which the trained data. 133 Convolutional neural networks feature among the
85 The problem led to the development of the deep134 most widely used deep learning techniques for
86 learning methodology. Deep learning is a subset135 image classification. Since CNNs learn directly
87 It has the propensity to learn in its own136 from input images, handcrafted feature extraction
88 computational approach. By using deep learning,137 is no longer required in this technique. This offers
89 a computer can identify words, sounds, and138 three benefits: sparsity of interactions, equal
90 images. To enable this learning, thousands and139 representations, and parameters shared.
91 thousands of images in datasets are used to train
92
 140 Another well-liked deep learning technique,168 vectors. To generate a structured prediction
141 particularly in voice and natural language169 across a variable size input structure, the same set
142 processing, is the recurrent neural network170 of weights is applied recursively over a structured
143 (RNN). RNN makes use of the network's171 input.
144 sequential data. Consider an RNN as a short-term
145 memory unit with input, output, and hidden172 A SIMPLE RvNNARCHITECTURE
146 layers (layers x, y).
147 Recursive Neural Network (RvNN) RvNN can173
148 categorize an item using compositional encoding174 We will now examine the use of Convolutional
149 and generate predictions in a hierarchical175 Neural Networks (CNNs) for picture
150 structure. 176 categorization.
151 177 A CNN consists of many layers.
152 178
153 The process begins with an input image. 179 1. The convolutional layer is responsible for
154 180 transmitting data to the following layer.
155 This generates a feature map through the181
156 application of several filters, subsequent182 2. The pooling layer pools a group of neurons to
157 application of ReLU functions to augment non-183 one neuron in the next level.
158 linearity, implementation of pooling layers for
184
159 each feature map, and transforming the images
185 3. A fully connected layer where each neuron in
160 into a singular elongated vector. The vector then
186 one layer is connected to every neuron of the
161 gets input in a fully connected artificial neural
162 network after flattening. Train the network by187 another layers.
163 using both forward and backward propagation.188 CNN uses many hidden layers, sometimes
164 Repeat this process until the network is well-189 hundreds of them, to enable feature recognition.
165 defined and has feature detectors and learned190 The characteristics learned are somewhat more
166 weights. 191 complex at each successive layer.
167 192 Various approaches to image identification
193 using CNN
194

195
196 Let's take a closer look at how CNN is used for226
197 picture categorization.

198 The first step in image classification is accepting

199 the incoming image and making sure it adheres to
200 the class description. them can be easily classified
201 by humans, but computers see them differently.
202
203 What we see
204
205

227
228
229
230
231 The output of the first convolution layer becomes
206 232 the input for the second layer as the picture passes
207 233 through it. Each and every convolution layer will
208 234 experience it.
209 235
210 What a computer see 236
237
211 Instead of seeing an image, a computer sees238 an The convolution layer is inserted before the
212 array of pixels; for example, if an image has 200239
by nonlinear layer. In the nonlinear layer, an
213 240
200 pixels, the array's size is 200 by 200 pixels, activation function is added, which provides with
214 with 200 being width, 200 for height, and 2241 for nonlinear properties; otherwise, without this
215 RGB values come 242
next. feature, a network cannot be intensive or describe
216 The computer searches for base level 243 the response variable.
217 characteristics, such as borders or curvature,244 in
218 order to solve these problems. 245 Next to the nonlinear layer comes the pooling
246 layer. Width and height of the image apply to the
219 The picture is initially sent into the convolution
247 execution of down sampling, thus reducing the
220 layer. The program then chooses a smaller matrix
248 volume of an image.
221 known as a filter. Next, the filter generates
222 convolution. Multiplying the filter's value by the
223 pixel value is its task. Next, the total of all these
224 multiplications is calculated.
225
249 Once the layers above are complete, a fully 280 Plotting this result will allow us to see that it shows
250 connected layer is added to which produces an281 N- 96% accuracy, which is a good result.
251 dimensional vector, where N is the number of
252 classes for which the model can pick to perform282
the
253 classification task. Adding the fully connected283 Literature review
254 layer does this.
284 Nurul Hazirah Indra and Mohd Azlan Abu (2019)
255 285 [1] used the TensorFlow framework to classify
256 To form a model, however important, it should 286
pass images using deep neural networks. Because
257 through various stages: 287 Python comes with TensorFlow, it was employed
258 288 in this manner. Approximately thousands of
259 Model building is actually started with the object
289 flowers were gathered, and five different varieties
260 model = Sequential(). 290 were present, each with hundreds of photos. All of
261 compose layers based on their model types.add291 the these data sets are trained by the DNN until the
262 type_layer function. 292 system recognizes every image, at which point
263 Once sufficient numbers of layers are included,293
the each image is classified. Their accuracy rate was
264 294
model is compiled. In addition to specifying an about 90%. Ram, S., & Chauhan, K. (2018)
265 optimization algorithm, a loss function will need to
295 [2]
266 be defined. Once developed, the model is applied
267 to a previously untouched data set and 296 its They used, in the article, a deep learning
268 effectiveness is measured. 297 convolutional network based on keras and
269 298 tesnsorflow with Python for the picture
270 After the training of the model is complete and299it categorization. Four different CNN structures on
271 produces the correct outcome, the model may300 be CPU systems were tested, as well as four different
272 301
saved.save("file_name"). The model applies in combinations of classifiers and activation systems
273 302
reality. This step is termed as model evaluation, - softmax, sigmoid classifiers, and Relu, Tanh
274 303
meaning that the data can be evaluated using the activation functions. They therefore defined
304 parameters for image classification in Python,
275 model.
305 then created a CNN with two convolutional
276 The model's accuracy after testing was 96%; the
306 layers, fitted the created CNN to the image
277 only issue was that it took the model around 50
307 dataset, trained it, and tested it and obtained
278 minutes to display the 60 epochs.
308 satisfactory accuracy. They then prepared an
279
309 image dataset with 10,000 images of dogs and
310 cats, of which 8,000 were used for training and
311 2000 for testing. Misbah Naimuddin Panchbhai
312 and Dr. Vinayak Bharadi (2017)
313

314
315 [3] have suggested deep learning approach known
316 as image classification, where the neural network
317 is used to classify the image.
318 One can use Python programming language to363 Their work makes a contribution by comparing
319 test how efficiently the system distinguishes364 the performance of several classification
320 between images on both the GPU and CPU. For365 algorithms with a deep learning approach for
321 this, the system will take a picture either through366 multi-class picture classification. Additionally, it
322 a digital camera or fetch one from a database. It367 contributes by demonstrating how easy it is to
323 will make adjustments to each picture in368 apply the deep learning method.
324 preparation for the next stage. Techniques for
325 dimensionality reduction include those of block369 .
326 transition coding and histograms among others,
327 feature extraction techniques. The authors used370 Feng-Ping An and Jun-e-Liu (2019) [6] Wrote
328 the CUDA library for the GPU and used Python371 the research called "Image Classification
329 and TensorFlow for the CPU. Their performance372 Algorithm based on Deep Learning – Kernel
330 evaluation covered a range of aspects such as373 Function". This way of computer vision includes
331 accuracy and the time taken to execute both for374 two tasks - first, the exact identification of
332 CPU and GPU. M. Neelima and M. Manoj375 pictures and then the extraction of features from
376 them. The course of action is introduced by
333 Krishna (2018).
377 utilizing the sparse representation of the deep
334
378 learning network architecture. The main
335 379 advantage of using the representation is that it is
336 [4] They used the architecture of AlexNet and the380 an effective way of solving the function
337 convolutional neural network for categorization381 approximation problem, which can be very
338 in order to classify pictures in the proposed382 complicated.
339 system. They selected four test photos from their383
340 database, resized the photos to different sizes384 Build an adaptive approximation capable model
341 corresponding to their respective section sizes385 with deep learning.
342 and ran the test. They educated the data in the386
343 network by providing a test image and told the387 Vijay Anand and M. MuthukrishnanRamprasath
344 network about the result of that particular image.388 (2018) [7] proposed the use of convolutional
345 And from this experiment, they saw that photos389 neural networks in classifying images. Toward
346 were rightly identified, even for the portion of390 computer vision objectives, the system uses the
347 test images it shows how deep learning391 deep learning algorithm. The system through the
348 algorithm is effective in image classification.
392 autonomous classification of images, makes use
349 W.N.A.Ismail, W. A. Ezat, and M. M.Desouky393 of CNN. Using it as a benchmark, MNIST
350 (2020) [5] In their study, the PASCAL VOC394 dataset enables the classification of gray-scale
351 dataset's pictures were classified using a CNN395 images. They achieved a result of 98% by
352 model that had been pre-trained on image-net.396 training using CNN network data. The classifier
353 The deep learning CNN model's performance397 training consists of four primary steps:
354 was enhanced by the application of the transfer398 preprocessing the data, feature extraction,
355 learning technique, which allowed for more399 normalization of the data set, and SVM image
356 accurate classification even with shorter400 classification. Xin Mingyuan and Wang Yong
357 computation times and less hardware. Results401 (2019).
358 from the SVM technique, the region ranking
359 SVM method, and the super vector coding of the
360 local image descriptor approach were compared
361 with the acquired results.
362
402 [8] proposed a research effort known as the441 fine-tuning, a hybrid, and the CNN model with
403 Deep Convolution Neural Network-Based Image442 pre-training. The CNN model was passed
404 Classification Model. The research into the443 through the network after pre-training and fine-
405 backpropagation method provided a444 tuning the model, whereas in the hybrid
406 foundational basis for the training of a deep445 approach, patch-based feature extraction is used.
407 neural network intended to reach the maximum446 These methods provide efficiency with high
408 span of minimal classification error. For this447 performance and speed as their prime
409 experiment, they explored three approaches: 448 advantages.
410
411 Published by Leonardo da Vinci and Sumanth Reference Advantages
412 Roychowdhary, a research paper on "Image Mohd Azlan 1- An accuracy
413 classification using deep learning and prior Abu, Nurul of 90% was
414 knowledge" was only present in their 2019 work. Hazirah found .
Indra(2019) [1]
415
416 [9] In their work, they brought historical Karan Chauhan, 1- A good
417 knowledge into deep learning methods by simply ShrwanRam(2018) [2] accuracy was
found in
418 using semantic-based regularization.
result.
419 2- Model took
420 This method can be utilized for group very small
421 classification or as part of a learning process. The amount of
422 CIFAR-10 dataset, which consists of 60,000 time
423 images, formed the basis of their exploratory fo
424 study. The training set was employed to train the r training.
425 models, while the test set provided results for Dr. Vinayak Bharadi, 1- The system
426 classification. The two primary benefits of their Misbah can measure
427 proposed approach were the very low costs NaimuddinPanchbhai( accuracy of
428 associated with test classification and the 2017) [3] classifying
429 impressively quick training of collective images on
430 knowledge—taking less than 6 minutes. CPU and
431 GPU.
432 M Manoj Krishna, M 1- Images were
433 Neelima(2018) [4] classified
434 The CNN model was passed through the network correctly for
435 after pre-training and fine-tuning the model, even portion
436 whereas in the hybrid approach, patch-based of test
437 feature extraction is used. These methods images.
438 provide efficiency with high performance and W. A. Ezat, M 1- It shows the
439 speed as their prime advantages MDesouky and N A simplicity of
440 Ismail(2020) [5] implementing deep
learning algorithm.
2- Comparison were
shown on the basis of
performance.
449 Jun-e-Liu, Feng-Ping 1- Approximatio
An(2019) [6] N problem of
450 complex
451 function was References:
452
453 solved.
454
Muthukrishnan 1- Good
Ramprasath, M. Vijay accuracy of
Anand(2018) [7] almost 98%
Mingyuan Xin 1- Better
and Yong spee
Wang(2019) [8] d and
performance
Michelangelo 1- Collective
Diligenti, classification
Soumal shows
iRoychowdhary(2019) improvement
[9] over
individual
classification
.
2- Cost of
collective
classification
was
negligible.
3- Error
reduction was
larger than
2.7%.
4- Time
consumption
was very less.
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486 Conclusion:
487 After examining several approaches, it can be said that each strategy may have a few benefits
488 over the others. For example, some strategies demonstrate more accuracy than others, while
489 some are more cost-effective. While there have been significant advancements in picture
490 categorization recently, there is always room for improvement.
491
492 1. Mohd Azlan Abu, Nurul HazirahIndra(2019) [1]- Image classification using deep neural
493 network.
494 2. Karan Chauhan, ShrwanRam(2018) [2]- Image classification using deep learning based
495 on keras and tensorflow.
496 3. Dr. Vinayak Bharadi, Misbah NaimuddinPanchbhai(2017) [3]- Image classification
497 using deep learning.
498 4. M Manoj Krishna, M Neelima(2018) [4]- Image classification using deep neural
499 network.
500 5. W. A. Ezat, M MDesouky and N A Ismail(2020) [5]- Image
501 classification using convolutional neural network.
502 6. Jun-e-Liu, Feng-Ping An(2019) [6] Image Classification Algorithm based on Deep
503 Learning – Kernel Function.
504 7. MuthukrishnanRamprasath, M. Vijay Anand(2018) [7]Image Classification using
505 Convolutional Neural Network.
506 8. Mingyuan Xin and Yong Wang(2019) [8] Image Classification Model based on Deep
507 Convolution Neural Network.
508 9. Michelangelo Diligenti,
509 SoumaliRoychowdhary(2019) [9] Image classification using deep learning and prior
510 knowledge.
511 10. https://in.mathworks.com/matlabcent ral/fileexchange/59133-neural- network-
512 toolbox-tm--model-for- alexnet-network
513 11. H. Lee, R. Grosse, R. Ranganath, and
514 A.Y. Ng. Convolutional deep belief networks for scalable unsupervised learning of
515 hierarchical representations.
516