LostNet: A smart way for lost and find

Meihua Zhou1, Ivan Fung2, Li Yang1, Nan Wan1, Keke Di1

Tingting Wang*
1
School of Medical Information, Wannan Medical Collage, Wuhu City, China

{mhzhou, 20120037, wannan, mhzhou0412}@wnmc.edu.cn

2
[email protected]

* School of Medical Information, Wannan Medical Collage, Wuhu City, China

[email protected]

Abstract:
Due to the enormous population growth of cities in recent years, objects are frequently lost
and unclaimed on public transportation, in restaurants, or any other public areas. While
services like Find My iPhone can easily identify lost electronic devices, more valuable
objects cannot be tracked in an intelligent manner, making it impossible for administrators to
reclaim a large number of lost and found items in a timely manner. We present a method that
significantly reduces the complexity of searching by comparing previous images of lost and
recovered things provided by the owner with photos taken when registered lost and found
items are received. In this research, we will primarily design a photo matching network by
combining the fine-tuning method of MobileNetv2 with CBAM Attention and using the
Internet framework to develop an online lost and found image identification system. Our
implementation gets a testing accuracy of 96.8% using only 665.12M GLFOPs and 3.5M
training parameters. It can recognize practice images and can be run on a regular laptop.

Keyword: artificial intelligence, lost and found, intelligent city

Introduction:
The population density and the quantity of lost objects are both rising in the areas where
urban rail transit is located around the world at the present time; yet, the traditional manual
search service is ineffective. In this situation, there is an immediate need to speed up the
development of intelligent lost and found systems in order to lessen the difficulty that
transportation operators face when it comes to lost and found management. Deep learning can
be used to build recognition and classification models for lost and found items. This is a new
method that can reduce reliance on manual labour, quickly and accurately identify categories,
significantly reduce human service costs for transportation operators, and better practice
green development strategies.
The application of convolutional neural networks is widespread throughout many scientific
subfields pertaining to image recognition and classification (Sun et al., 2021). It is not a
passing fad that academics in a wide variety of fields have shifted their attention to the study
and practical use of image recognition. In the realm of trash classification, a method for
garbage image classification was devised. It was built on an enhanced version of MobileNet
v2 and paired with transfer learning to increase the real-time performance and accuracy of
garbage image classification models (Huang et al., 2021).

Research Background:
Many academics are also engaged in questioning the status quo in the many subfields that
make up the area of applied image recognition. (Yang et al., 2015) attempting to identify
plant leaves by the utilization of a hierarchical model that is based on CNN. A study on
establishing the optimal size of the training data set that is required to achieve high
classification accuracy with low variance in medical image classification systems is presented
by Cho et al. (2015). (Purnama et al., 2019) offer a method for the classification and
diagnosis of skin diseases that is suitable for use in teledermatology.

It has also been demonstrated that transfer learning is beneficial in a variety of contexts.
Convolutional neural networks are used in the methodology that Lee et al. (2016) propose as
a fine-grained classification method for large-scale plankton databases. The implementation
of transfer learning in CNN is one potential solution. (Liu et al., 2020) apply unsupervised
transfer learning to CNN training to address these problems. Specifically, they transform
similarity learning into deep ordinal classification with the assistance of several CNN experts
who were pretrained over large-scale-labelled everyday image sets. These CNN experts
jointly determine image similarities and provide pseudo labels for classification. (Purwar et
al., 2020) make use of some models that are related to convolutional neural networks (CNNs)
in order to identify mesangial hypercellularity in MEST-C. (Herzog et al., 2021) concentrate
on the classification of MRI for the diagnosis of early and progressive dementia by utilizing
transfer learning architectures that employ Convolutional Neural Networks-CNNs, as a base
model, and fully connected layers of SoftMax functions or Support Vector Machines-SVMs.
(Phankokkruad, 2021) propose the three CNN models for detecting lung cancer using
VGG16, ResNet50V2, and DenseNet201 architectures. The proposed method is based on
transfer learning.

Methodology:
In a procedure in which CNN is being used to address more and more components of the
problem, the lost and found problem has not been solved in any intelligent method as of yet.
As a result, we recommend taking a methodical approach by utilizing Mobilnet v2 and an
intuitive graphical user interface (GUI). In this study, we build on earlier research to further
investigate the detection and categorization of lost and discovered items, and we present an
approach that combines perceptual hashing with MobileNet v2 transfer learning.
In order to solve the relatively complex problems associated with the image dataset of lost
and found items, we carried out extensive research, classified the most common items that are
lost and found into ten categories using questionnaires and market research, and produced
private dataset images using crawlers, real-world photography, and research examples. After
that, the generated dataset is used to train the network and establish an intelligent recognition
classification model for lost and found images. This model solves the problems of labor cost
and time cost consumption that are associated with conventional methods and proposes an
accurate and complete solution with scientific and accurate experimental data.
I. Mobilenet v2
An outstanding example of a good lightweight convolutional neural network is MobileNetV2
(Sandler et al., 2018). The network creates a reverse residual and linear bottlenecks, both of
which are helpful for feature extraction; the linear activation that is used in the final layer of
the inverted residual structure prevents the loss of low-latitude information; and the
traditional convolution is replaced by depth-separable convolution, which significantly
reduces the amount of the model's calculations as well as the number of the parameters that it
uses. It was developed specifically for pictures and has applications in image categorization
as well as the development of generic features.
Convolution on a depth-wise and point-wise scale are the two components that make up
depth-separable convolution. The depth separable convolution algorithm is not the same as
the standard convolution algorithm. During the process of convolution, each channel of the
feature map is covered by exactly one convolution nucleus, and the total number of
convolution nuclei is equal to the total number of channels. The following phrases can be
used to describe depth convolution:

𝑂𝑥,𝑦,𝑐 = ∑𝑊,𝐻
𝑤,ℎ
𝐾𝑤,ℎ,𝑐 ⋅ 𝐼𝑥+𝑤,𝑦+ℎ,𝑐 （formula1）

In the equation presented above, the variable O stands for the output feature graph, c stands
for the channel of the feature graph, x and y stand for the coordinates of the output feature
graph on channel, K stands for the convolutional kernel with wide W and high H, I stand for
the input feature graph, and w and h stand for the convolutional kernel weight element
coordinates of channel.
The primary difference between point-by-point and standard convolution is the size of the
convolution kernel, which in point-by-point convolution is fixed at 1x1. The first step of
depth separable convolution is to employ depth convolution to extract the characteristics of
each channel. Next, point-by-point convolution is used to correlate the extracted channel
characteristics. The depth separable convolution is intended to reduce the number of
parameters and computations required by the conventional convolution. When compared to
the total number of calculations involved in the traditional convolution, the results are as
follows:
𝑅1 𝐷𝑓 2 𝐷𝑘 2 𝐼+𝐷𝑓 2 𝐼𝑂 1 1
= =𝑁+ （formula2）
𝑅2 𝐷𝑓 2 𝐷𝑘 2 𝐼𝑂 𝐷𝑘 2

In the formula 2 shown above: R1 and R2 represent the calculations of depth separable
convolution and standard convolution, respectively; Df and Dk represent the height and width
of the input feature matrix; I represent the depth of the input feature matrix; and O represents
the depth of the output feature matrix.
In the process of feature extraction, MobileNetV2 makes use of a depth detachable
convolution with a size of 3x3, which means that the calculation cost is 1/9 of what it would
be for a standard convolution. However, the reduction in accuracy is very minimal, which is
also one of the most notable qualities of the model.
Figure 1 presents the organizational structure of the MobileNetV2 network. It is primarily
made up of three components: the front end is a convolutional neural network (CNN), which
is constructed by several layers of convolution, and then the average pooling of 1,280 7x7
blocks is utilized to generate 1,280 nerves. Element, which was afterwards completely
coupled with one thousand neurons.

Figure1 MobileNet v2 based classifier

MobileNet v2 borrows ResNet's residual connection notion and uses it to suggest a reverse
residual structure, as seen in the picture. The structure employs PW convolution to increase
dimension, 3x3 DW convolution to extract channel characteristics, then PW convolution once
again to decrease feature dimension. Notably, as seen in Figure 2, a residual link exists only
when the step duration is 1. In the second situation, when the step size is 2, the series
structure is employed directly, and the reverse residual structure is upgraded and then
lowered, so that the network permits smaller input and output dimensions, hence decreasing
the amount of computation and parameters of the network. Simultaneously, residual
connection may increase the effectiveness of gradient propagation and deepen the network
layer.

Figure2 illustrative example in graphical form of the MobileNetV2 architecture.

MobileNet v2 networks are well-suited for implementation on mobile terminals or embedded
systems because to their small model size, cheap processing capacity, and high computation
speed in comparison to more conventional convolutional neural networks. Moreover, it can
keep up with the CPU's strict speed needs.
II. Transfer learning
It takes a significant amount of time and resources to train a deep neural network model on ea
ch job from the very beginning. The learning process is laborious and time consuming. In ord
er for the network to begin its learning process, it must first texture the outline. Subsequently,
the network must extract the characteristics after increasing the number of network layers. H
owever, in the majority of instances, academics will use the information that they have gained
in the past to apply to the current challenges that they face. Learning by transfer is the name
given to this kind of information transmission (Zhuang et al., 2021). In the field of deep learn
ing, transfer learning is combined with it, and the pre-training weight obtained by the model i
n similar tasks is used. This is done to avoid model training from scratch, reduce the cost con
sumption caused by the network model due to new learning, accelerate the network converge
nce speed, and further improve the stability and generalization ability of the model (Song et a
l., 2021).
The ImageNet data collection includes one thousand categories and encompasses one million
two hundred thousand pictures. There is no shadow of a doubt that the utilization of pre-train
ed network models on such a massive data collection is capable of being successfully ported t
o a variety of picture categorization endeavors.
In this study, transfer learning is carried out by pre-training a deep CNN model using a portio
n of the ImageNet dataset. This portion of the dataset contains image classification data. A cl
assifier was added to the pre-trained Mobilenet v2 model, and then it was used to download t
he private dataset and split it into 10 primary categories. The pre-trained model was imported
via making use of the torch vision library, and samples from the private dataset were used for
the training and testing of the model. By doing things this way, we were able to cut down on t
he time spent training the features extractor, which is what takes the most time for the majorit
y of models, and instead train the classifier directly. Because of this, the amount of money sp
ent on training has decreased.

Figure3 transfer learning

 III. Convolutional block attention module
The Convolutional Block Attention Module, often known as CBAM, is a type of attention
mechanism module that integrates Channel and Spatial. This module may be embedded into
the Model module, trained end-to-end with Model, and only contributes a little amount of
additional processing (Woo et al, 2018). Figure 4 demonstrates the Channel attention module
in addition to the Spatial attention module. The Channel Attention Module is capable of
taking an Intermediate Feature Map as an input and outputting both a 1-D Channel Attention
Map and a 2-D Spatial Attention Map through the use of the CBAM. In order to aggregate
the spatial information, the average pooling and maximum pooling methods are utilized. The
approaches provide two descriptors, which are then passed to the same shared network in
order to produce the channel attention map. The complete channel attention map can then be
acquired by doing the following:
𝑀𝑐 (𝐹) = 𝜎 (𝑀(𝑎(𝐹)) + 𝑀(𝑚(𝐹))) = 𝜎 (𝑊1 (𝑊0 (𝐹avg )) + 𝑊1 (𝑊0 (𝐹max ))) (formula

3)
The spatial attention module can highlight the information region and generate two 2D maps,
which are then linked and convolved by a standard convolutional layer to produce a 2D
spatial attention map. This map is produced after first averaging and maximum pooling along
the channel to generate efficient feature descriptors. In conclusion, the formulation for the
output of the spatial attention mapping is as follows:
𝑀𝑠 (𝐹) = 𝜎 (𝑓 (𝑓𝑐 (𝐹avg ，𝐹max ))) (formula4)

F is the input feature graph, is the sig-moid nonlinear activation function, M is the forward
calculation function of multi-layer perceptron without bias, a and m are the mean and
maximum pooling functions respectively, W0 and W1 are the weights of 2 linear layer, and
Favg and Fmax are the mean and maximum pooling functions respectively. In the equation
that was just presented, F represents the input feature graph, represents the sig-moid
nonlinear.

Figure4 Channel attention module and Spatial attention module

It is necessary to arrange the Channel attention module and the Spatial attention module in a
sequential fashion in order to generate the CBAM. This provides the model with the ability to
concentrate on important features in both the channel and spatial dimensions while
suppressing features that are not necessary. Figure 5 provides a high-level view of CBAM,
while the following provides a concise summary of the complete calculating process:
 𝐹 ′ = 𝑀𝑐 (𝐹) ⊗ 𝐹 (formula5)
𝐹 ″ = 𝑀𝑠 (𝐹 ′ ) ⊗ 𝐹 (formula6)
It is necessary to arrange the Channel attention module and the Spatial attention module in a
sequential fashion in order to generate the CBAM. This provides the model with the ability to
concentrate on important features in both the channel and spatial dimensions while
suppressing features that are not necessary. Figure 5 provides a high-level view of CBAM,
while the following provides a concise summary of the complete calculating process:

Figure5 The overview of CBAM

This algorithm integrates the attention mechanism in the dimension of the channel and the
space, strengthens the important features while suppressing the unimportant features, and
further highlights the required features in order to improve the classification accuracy of the
model. The CBAM structure is introduced in the first layer of the Mobilenet v2 network by
this algorithm. Because CBAM is a lightweight module, it does not have a significant impact
on the total number of model parameters, but it does improve emphasis on the model's
primary features. This effectively guarantees that the benefits of a lightweight model will be
realized.

IV. Perceptual hash algorithm

The perceptual hash algorithm (Samanta & Jain, 2021) begins by reducing the size of the
picture and simplifying the colors. Next, it aggregates the decomposition frequency and
trapezoidal shape of the picture by using the discrete cosine transformation method. After
that, it reduces the DCT, keeps the 8*8 matrix in the upper right corner, calculates the
average value of all 64 values, further reduces the DCT, and finally calculates the hash value.
The algorithm creates a unique string that acts as a "fingerprint" for each image, and it then
analyzes the results by contrasting the various "fingerprint" strings. When the result is closer,
it indicates that the picture is more comparable. The advantage of using this approach is that
regardless of whether the height, width, brightness, or color of the image is modified, the
result value of the hash value will not remain intact. This allows one to prevent the impact
that would be caused by the modest modification of the image. An online image
identification system is constructed in this research using the perceptual hash method and the
concept of searching for images using photos.
V. Measurement
There is currently no standardized dedicated data collection that can be used for the
categorization of photos of lost and found items. The study of confidential data sets is the
focus of this work. Using the results of the poll, streamline the complicated lost and found
categories into the top 10 categories that are the easiest to lose track of, with a total of 10499
photographs. In addition, in order to match the conditions of a practical application, we have
not performed any pre-processing on the dataset. Directly scaling the collected lost and found
data set pictures to the input size of the network model will directly lead to the loss of some
information in the picture, or even direct distortion, and will ultimately directly affect the
image classification effect and recognition accuracy. This is because the original image had
an uneven quality and a high resolution. As a result, prior to the pre-training, this article
automates the data set photos in batches by using the design action in Adobe Photoshop
software. Additionally, the article standardizes the image resolution to 300 x 300, which
helps to minimize the overall picture size. The data processing of the sample data has to be
improved in this work in order to increase the capacity of the model to generalize its findings.
The 10,499 common lost and lost photographs augmented by the data were divided into the
training set and validation set with 70% and validation set with 30%.
Table 1. Sample distribution of the lost object Figure slice dataset used in
this paper

Class name Number of images

bag 1036
book 1067
card 1043
earphone 1025
key 1070
lipstick 1066
Phone 1068
umbrella 1045
USBflashdisk 1020

vacuumcup 1059

Figure6 The data inside the dataset

This article employs Accuracy, AgePrecision (AP), Recall, Precision, and Loss as model
performance evaluation indicators in order to analyse the advantages and disadvantages of the
model.
Calculating values using typical complex matrix approaches like as accuracy, recall, and
precision may be difficult. When it comes to image classification, the confusion matrix is
most commonly employed to make a comparison between the classification and the actual
measurement value. This allows for a more understandable and accurate description of the
correctness of model categorization. The formula for the computation is as follows:
𝑇𝑁+𝑇𝑃
Accuracy = 𝑇𝑁+𝐹𝑁+𝐹𝑃+𝑇𝑃
（formula7）
𝑇𝑃
Recall =
𝑇𝑃+𝐹𝑁
（formula8）
𝑇𝑃
𝑃recision = 𝑇𝑃+𝐹𝑃
（formula9）
The number of successfully predicted positive samples is denoted by the letter TP in the
formula 7,8,9, the number of correctly anticipated negative samples is denoted by the letter
TN. FN is for the number of samples that had errors predicted for them when they were
positive, and FP stands for the number of samples that had errors predicted for them when
they were negative.
The performance of the model recognition accuracy of the selected model may be evaluated
in an objective manner by using the average accuracy rate. The mathematical formula for the
calculation is as follows:
𝐸
∑ Accuracy
1
𝐴𝑃 = 𝐸
（formula10）
The term AP, which is utilized in the aforementioned formula 10, stands for the average
accuracy of the model that was picked, while E stands for the total number of iterations, and
Accuracy stands for the accuracy of each iteration of the model that was selected.
The degree to which the model's predicted value and the actual value deviate from one
another may be estimated using the loss value. If the loss value is lower, then the model's
predicted result will be closer to the actual outcome. The formulae for the calculations are as
follows:
1 1
𝐿𝑂𝑆𝑆 = 𝑁 ∑𝑖 𝐿𝑖 = 𝑁 ∑𝑖 − ∑𝐾 𝑦 𝑙𝑜𝑔( 𝑝𝑖𝑐 )
𝑐=1 𝑖𝑐
（formula11）
In the previous formula number 11, Loss is a numerical representation of the loss value
associated with the model that was chosen, and K is the total number of categories. pic is the
anticipated probability that sample I belongs to category c. If the category is the same as the
category of sample I it has a value of 1, and otherwise, it has a value of 0. The cross-entropy
loss function is a convex function, and it is possible to calculate the value that is best for the
whole world.
Experimental results:
I. Training
In the course of our research, we looked at the problem of lost-object picture categorization.
In order to classify images, we made use of Convolutional block attention module in
conjunction with the well-known framework MobileNet v2. We utilized this architecture
(transfer learning) to speed up the learning process while also reducing the amount of time
needed for training. In order to evaluate the effectiveness of transfer deep learning, we
examined the outcomes of three different optimizer functions: the Adaptive Moment
Estimation Algorithm (ADAM), the Root Mean Square Propagation (RMSprop), and the
Stochastic gradient descent (SGD) algorithm. We noticed that the optimizer SGD function
functioned in a better method and got an accuracy of 96.8 percent, while the loss detected
was 0.0047. We figured this out by looking at the table. The results of the SGD function's
performance are presented in the form of a linear graph of the accuracy and loss parameters
in the figure. When compared to the other optimizers, SGD's training time for the model was
significantly shorter, despite the fact that all other model parameters remained same.

Table2. Comparative chart of results using optimizer function

Components SGD RMSprop ADAM

Accuracy/% 96.8 89.5 94.7
Loss 0.0047 0.0050 0.0048
Freeze Epoch 50 50 50
Freeze batch size 32 32 32
UnFreeze Epoch 400 400 400
Unfreeze batch size 64 64 64
Init lr 1e-2 1e-2 1e-2
Lr decay type cos cos cos
momentum 0.9 0.9 0.9
Execution Time/ minutes 5.38 14.28 10.71

Figure7a(left) Training and validation curve of the model

Figure7b(right) the valuation index of different measure index

It is abundantly evident, after reviewing the preceding explanation complete with graphs and
table, that the optimizer function SGD performed far better than the performance of the other
functions when applied to the issue of lost-object picture classification.
EfficientNet, Inception v4, ViT-B. 32, DenseNet201, VGGNet19, ShuffleNet v2, ResNet152
and MobileNet v3 were the eight transfer learning algorithm models of the same kind that
were chosen at random for the purpose of doing a comparison study in order to test the
efficiency and superiority of the research approach that was provided in this work.The
optimal parameter values of the model convergence were selected as the respective
parameters.This study's private data set is used to train and verify a total of nine models using
these settings. The outcomes of the training are presented in Table 3.
Table3. Results of the selected model training

Accura Total
Model AP/%
cy/% parameters/M
EfficientNet
93.9 89.0 66.348
（Koonce et al，2021）
Inception v4
89.5 80.3 41.158
（Szegedy et al，2017）
ViT-B/32
（Dosovitskiy et al， 94.9 94.3 104.766
2010）
DenseNet201
（Huang et al， 95.9 95.4 20.014
2017）
VGGNet19
（SIMONYAN et al， 95.6 94.7 139.611
2021）
ShuffleNet v2
88.6 82.3 2.279
（Ma et al，2018）
ResNet152
87.8 81.9 60.193
（He et al，2016）
MobileNet v3
95.7 94.9 5.483
（Koonce et al，2021）
us 96.8 96.2 3.505

Figure8 The performance, accuracy and parameter amount compare of difference models
The algorithm that is suggested in this article is transfer learning, which is based on the
improved MobileNetv2. This can be seen in Table 3. The results of the experiments
demonstrate that the method under study has an average accuracy of 96.2% when applied to
data sets that were self-built. This is a higher accuracy rate than EfficientNet, Inception v4,
ViT-B. 32, DenseNet201, VGGNet19, ShuffleNet v2, ResNet152 and MobileNet v3
correspondingly. 7.2%, 15.9%, 1.9%,0.8%,1.5%, 13.9%, 14.3% and 1.3% of the total
learning was transferred from other models. The suggested technique has obtained the
greatest accuracy rate in private data sets, which is 96.8%, along with strong generalization
ability and resilience. When compared to the similar kind of transfer learning model that has
been developed, this is how it stands out. In correlation with the information shown in Figure
8, The ordinate in the figure shows the test accuracy, the abscissa represents GFLOPs, the
circle colors reflect distinct transfer learning models, and the figure size denotes total
parameters. The suggested method gets the maximum accuracy on private datasets, and it is
robust and has a decent ability to generalize. When Figure 8 and Table 3 are considered
together, a further complete evaluation of the performance of each model is carried out. This
model is only second to MobileNet v3 and ShuffleNet v2, but it is worth recognizing that the
Total parameters of the proposed model are almost half that of MobileNet v3 Total
parameters, and the accuracy is higher than ShuffleNet v2 8.2%. MobileNet v3 and
ShuffleNet v2 are the only models that are ahead of this model. In comprehensive
comparison, the model that is being offered is a fantastic lightweight network, which
provides the possibility that the model might be transplanted to mobile devices.

Figure9a(left) Confusion matrix of the model

Figure9b(right) ROC curve of the model
We decided to combine the confusion matrix from the model single-picture test with the
ROC curve for the study so that we could have a more in-depth look at the data. The
confusion matrix is the method for measuring the accuracy of classification models that is the
simplest and most straightforward. Calculating the confusion matrix of the suggested model
requires using equation (7) and equation (9). This matrix represents the identification result of
the different types of lost object photographs and is derived from these equations. The
confusion matrix is shown in Figure 9a, with the rows representing the expected category for
the item that was lost, and the columns representing the actual category for the object that
was lost. Figure 9b illustrates the characteristic curves of the class 10 lost item categories for
the enhanced MobileNetv2 transfer learning model. ROC stands for receiver operating
characteristic. The fact that the curves for all ten classes are located in the top left corner is
evidence that the model is accurate, and the fact that the area under the ROC curve for each
class is equal to one further demonstrates the superiority and efficiency of the model that was
provided.
 II. Inference
In order to successfully apply the model to the actual world, it needs to be simple enough that
it can be executed on most laptops. In order to accomplish this, we implement this model as
well as a model that is very close to it on a regular computer that has a CPU, and we
comprehensively consider the inference speed of the computer as the evaluation index. It
takes the model 1.5 seconds to reason per picture on the CPU, which is significantly longer
than the standard migration model of this type or even 1/16 of the EfficientNet's processing
time. This can be seen in Figure 10a. Because it is such an exceptionally lightweight network,
it is not hard to imagine that the concept will eventually be adapted to work with other mobile
devices in the not too distant future. In light of this, we provide an engineering system that
integrates the perceptual hashing method.

This paper, which is based on the model training using the pytorch and torchvision machine
learning frameworks, uses the Spring Boot framework as the back-end service of the web
page in order to realize the online recognition of pictures. The front-end web page uses the
Layui framework in order to port the trained model to the Web end in order to realize the
online recognition of images that are uploaded by users. The steps involved in the online
identification procedure are as follows: the user launches the browser, navigates to the
system's website, navigates to the system's home page, clicks the button to upload the local
image, the front end sends the POST request to the back end, and the image is transmitted to
the back end. After the back end has received the picture that was uploaded by the user, it
will first make a request to the trained and upgraded MobileNet v2 migration learning model
for prediction and recognition, and it will then return a category. After it has determined the
category, it will next submit a request to the database, asking it to deliver an array containing
the picture address associated with the category it has just determined. After obtaining the
address, the image is then downloaded by using the image address, and then the perceptual
hash algorithm is used to compare the downloaded image to the image that was uploaded by
the user. The alignment will return a number for the similarity error, and the array will be
used to send the few photos that have the least significant value for the similarity error to the
front end. Figure 10b depicts the engineering interface after the model has been applied to the
situation.

Figure10a(left) The speed of inference of each model in cpu

Figure10b(right) The improved model is applied to the actual engineering interface diagram
Concludes:
This research presents a novel design plan in light of the difficulties that are associated
with the management of urban traffic operators. Even while the method of merging
convolutional neural networks with transfer learning has caused quite a stir in the world of
image recognition, it is not widely utilized when it comes to the automatic recognition of
photos that have been misplaced or found. This study offers a new technique of intelligent
picture identification based on hash algorithm to give full play to the advantages of hash
algorithm. The approach makes use of the concept of "search by map," and it is based on
hash algorithm. Because there are a significant number of missing images in the library and a
great number of feature points that have been extracted, the CBAM structure has been
introduced in the first layer of the network. This has been done in order to integrate the
attention mechanism in the dimension of channel and space in order to further highlight the
required features, which will ultimately lead to an improvement in the classification accuracy
of the model. Improved MobileNetv2 is used to establish a transfer learning training model to
identify common lost objects. This is done in order to reduce the number of ways in which
lost objects and database pictures can be compared to one another, thereby making it possible
to search for objects in a more convenient manner. Extensive experimental arguments on
photos taken from the loss-object dataset, employing a variety of transfer learning models.
The findings indicate that the proposed model has a high recognition accuracy in the GFLOPs
and Total parameters extremes, and the features derived by this model significantly beat those
extracted by the other approaches when it comes to the classification job. In order to enhance
the precision with which missing object categories may be identified, the proposed model is
being developed and put into practice in the search sector.
However, further solutions are required in order to further increase the accuracy of the
model and the segmentation of the various kinds of lost object. In order to do this, the
following stage will involve further subdividing each category so that it can be utilized in a
wider variety of search circumstances.

Acknowledgement:
This work was supported by Provincial College Students' Innovation and Entrepreneurship
Project Project for College Students [Grant numbers S202110368112]; University
Humanities and Social Science Research Program of Anhui Province [Grant numbers
SK2020A0380]; School level Project of Key Humanities and Social Sciences Research Base
of Anhui Province, Center for Mental Health Education of College Students [Grant numbers
SJD202001]; and School level Project of the Young and Middle-Aged Natural Science
Foundation of Wannan Medical College[Grant numbers WK202115]

Reference:

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep

convolutional Neural Networks,” Communications of the ACM, vol. 60, no. 6, pp. 84–
90, 2017.
D.-S. Huang, K.-H. Jo, J. Li, V. Gribova, V. Bevilacqua, and A. Hussain, Intelligent
computing theories and application 17th International Conference, ICIC 2021,
Shenzhen, China, August 12-15, 2021: Proceedings. Cham, Switchland: Springer,
2021.

F. Zhuang, Z. Qi, K. Duan, D. Xi, Y. Zhu, H. Zhu, H. Xiong, and Q. He, “A comprehensive
survey on Transfer Learning,” Proceedings of the IEEE, vol. 109, no. 1, pp. 43–76,
2021.

G.-W. Yang and H.-F. Jing, “Multiple convolutional neural network for feature extraction,”
Intelligent Computing Theories and Methodologies, pp. 104–114, 2015.

H. Lee, M. Park, and J. Kim, “Plankton classification on imbalanced large scale database via
convolutional neural networks with transfer learning,” 2016 IEEE International
Conference on Image Processing (ICIP), 2016.

I. K. Purnama, A. K. Hernanda, A. A. Ratna, I. Nurtanio, A. N. Hidayati, M. H. Purnomo, S.

M. Nugroho, and R. F. Rachmadi, “Disease classification based on dermoscopic skin
images using convolutional neural network in Teledermatology System,” 2019
International Conference on Computer Engineering, Network, and Intelligent
Multimedia (CENIM), 2019.

J. Cho, K. Lee, E. Shin, G. Choy, and S. Do, “How much data is needed to train a medical
image deep learning system to achieve necessary high accuracy?,” arXiv.org, 07-Jan-
2016. [Online]. Available: https://arxiv.org/abs/1511.06348. [Accessed: 20-Aug-2022].

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,”
arXiv.org, 10-Dec-2015. [Online]. Available: https://arxiv.org/abs/1512.03385.
[Accessed: 20-Aug-2022].

K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image
recognition,” arXiv.org, 10-Apr-2015. [Online]. Available:
https://arxiv.org/abs/1409.1556. [Accessed: 20-Aug-2022].

M. Phankokkruad, “Ensemble transfer learning for lung cancer detection,” 2021 4th
International Conference on Data Science and Information Technology, 2021.

M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted

residuals and linear bottlenecks,” 2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition, 2018.

M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted

residuals and linear bottlenecks,” arXiv.org, 21-Mar-2019. [Online]. Available:
https://arxiv.org/abs/1801.04381. [Accessed: 20-Aug-2022].

N. J. Herzog and G. D. Magoulas, “Deep learning of brain asymmetry images and transfer
learning for early diagnosis of dementia,” Proceedings of the International Neural
Networks Society, pp. 57–70, 2021.
P. Samanta and S. Jain, “Analysis of perceptual hashing algorithms in image manipulation
detection,” Procedia Computer Science, vol. 185, pp. 203–212, 2021.

S. Purwar, R. Tripathi, A. W. Barwad, and A. K. Dinda, “Detection of mesangial

hypercellularity of Mest-C score in immunoglobulin A-nephropathy using deep
convolutional neural network,” Multimedia Tools and Applications, vol. 79, no. 37-38,
pp. 27683–27703, 2020.

W. Song, Y. Wang, and Wang, Image and graphics technologies and applications.
Singapore: Springer Singapore, 2021.

Y. Liu, L. Ding, C. Chen, and Y. Liu, “Similarity-based unsupervised deep transfer learning
for remote sensing image retrieval,” IEEE Transactions on Geoscience and Remote
Sensing, vol. 58, no. 11, pp. 7872–7889, 2020.

Z. Sun, X. Yuan, X. Fu, F. Zhou, and C. Zhang, “Multi-scale capsule attention network and
joint distributed optimal transport for bearing fault diagnosis under different working
loads,” Sensors, vol. 21, no. 19, p. 6696, 2021.

Woo S, Park J, Lee J Y, et al. Cbam: Convolutional block attention module[C]//

Proceedings of the European conference on computer vision (ECCV). 2018: 3-19.

Weng L, Preneel B. A secure perceptual hash algorithm for image content

authentication[C]//IFIP International Conference on Communications and Multimedia
Security. Springer, Berlin, Heidelberg, 2011: 108-121.

Koonce B. EfficientNet[M]//Convolutional neural networks with swift for tensorflow.

Apress, Berkeley, CA, 2021: 109-123.

Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, inception-resnet and the impact of

residual connections on learning[C]//Thirty-first AAAI conference on artificial
intelligence. 2017.

Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers

for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020.

Huang, G., Liu, Z., Van Der Maaten, L. & Weinberger, K. Q. Densely connected
convolutional networks. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition. 4700–4708 (2017).

SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image

recognition [EB/OL]. [2021-02- 22]. http://arxiv.org/abs/1409.1556.

Ma N, Zhang X, Zheng H T, et al. Shufflenet v2: Practical guidelines for efficient cnn
architecture design[C]//Proceedings of the European conference on computer vision
(ECCV). 2018: 116-131.

He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//

Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-
778.

Koonce B. MobileNetV3[M]//Convolutional Neural Networks with Swift for Tensorflow.

Apress, Berkeley, CA, 2021: 125-144.