VGG16 Network Research Articles

A recurrent network based on active learning for the assessment of fish feeding status

The image-based fish feeding status analysis has the potential to bring the aquaculture industry a more intelligent and precise feeding process. However, most of this type of research currently is predicated on a fully supervised approach that not only requires experienced experts for sample labeling, but also the process appears to be time consuming and resource intensive. This makes it very difficult to obtain labeled samples. To solve this problem, we proposed a model method that combines the VGG16 network with active learning to reduce the number of labeled samples needed in the analysis of fish feeding status. The algorithm is composed of two modules: feature extraction classifier and information extractor. In addition, multiple training strategies such as Mish activation function and RAdam optimizer were used in the feature extraction classifier to improve the classification performance. Optimal selection of information entropy method was used to obtain the most informative samples. To evaluate the effectiveness of our method, we designed a fish feeding experiment and verified the performance of the proposed algorithm with a special focus on the accuracy and the number of labeled samples required. The results showed that the proposed algorithm has achieved satisfying results in accuracy (the test accuracy is 98%) and the number of labeled samples (approximately 1/10 of the original dataset). Therefore, the method proposed in this paper can be applied to aquaculture production to provide a reference for farmers' feeding operations.

Evaluation of Online Teaching Quality Based on Facial Expression Recognition

In 21st-century society, with the rapid development of information technology, the scientific and technological strength of all walks of life is increasing, and the field of education has also begun to introduce high and new technologies gradually. Affected by the epidemic, online teaching has been implemented all over the country, forming an education model of “dual integration” of online and offline teaching. However, the disadvantages of online teaching are also very obvious; that is, teachers cannot understand the students’ listening status in real-time. Therefore, our study adopts automatic face detection and expression recognition based on a deep learning framework and other related technologies to solve this problem, and it designs an analysis system of students’ class concentration based on expression recognition. The students’ class concentration analysis system can help teachers detect students’ class concentration and improve the efficiency of class evaluation. In this system, OpenCV is used to call the camera to collect the students’ listening status in real-time, and the MTCNN algorithm is used to detect the face of the video to frame the location of the student’s face image. Finally, the obtained face image is used for real-time expression recognition by using the VGG16 network added with ECANet, and the students’ emotions in class are obtained. The experimental results show that the method in our study can more accurately identify students’ emotions in class and carry out a teaching effect evaluation, which has certain application value in intelligent education fields, such as the smart classroom and distance learning. For example, a teaching evaluation module can be added to the teaching software, and teachers can know the listening emotions of each student in class while lecturing.

Optimizing the Deep Neural Networks by Layer-Wise Refined Pruning and the Acceleration on FPGA.

To accelerate the practical applications of artificial intelligence, this paper proposes a high efficient layer-wise refined pruning method for deep neural networks at the software level and accelerates the inference process at the hardware level on a field-programmable gate array (FPGA). The refined pruning operation is based on the channel-wise importance indexes of each layer and the layer-wise input sparsity of convolutional layers. The method utilizes the characteristics of the native networks without introducing any extra workloads to the training phase. In addition, the operation is easy to be extended to various state-of-the-art deep neural networks. The effectiveness of the method is verified on ResNet architecture and VGG networks in terms of dataset CIFAR10, CIFAR100, and ImageNet100. Experimental results show that in terms of ResNet50 on CIFAR10 and ResNet101 on CIFAR100, more than 85% of parameters and Floating-Point Operations are pruned with only 0.35% and 0.40% accuracy loss, respectively. As for the VGG network, 87.05% of parameters and 75.78% of Floating-Point Operations are pruned with only 0.74% accuracy loss for VGG13BN on CIFAR10. Furthermore, we accelerate the networks at the hardware level on the FPGA platform by utilizing the tool Vitis AI. For two threads mode in FPGA, the throughput/fps of the pruned VGG13BN and ResNet101 achieves 151.99 fps and 124.31 fps, respectively, and the pruned networks achieve about 4.3× and 1.8× speed up for VGG13BN and ResNet101, respectively, compared with the original networks on FPGA.

Comparison of Visual Features for Image-Based Visibility Detection

Abstract The image-based visibility detection methods have been one of the active research issues in surface meteorological observation. The visual feature extraction is the basis of these methods, and its effectiveness has become a key factor in accurately estimating visibility. In this study, we compare and analyze the effectiveness of various visual features in visibility detection from three aspects, namely, visibility sensitivity, environmental variables robustness, and object depth sensitivity in multiscene, including three traditional visual features such as local binary patterns (LBP), histograms of oriented gradients (HOG), and contrast as well as three deep learned features extracted from the Neural Image Assessment (NIMA) and VGG-16 networks. Then the support vector regression (SVR) models, which are used to map visual features to visibility, are also trained, respectively based on the region of interest (ROI) and the whole image of each scene. The experiment results show that compared to traditional visual features, deep learned features exhibit better performance in both feature analysis and model training. In particular, NIMA, with lower dimensionality, achieves the best fitting effect and therefore is found to show good application prospects in visibility detection. Significance Statement The visual feature extraction is a basic step for image-based visibility detection and significantly affects the detection performance. In this paper, we compare six candidate visual features, including traditional and deep learned features, from visibility sensitivity, environmental variables robustness, and object depth sensitivity in multiscene. Then the SVR models are also trained to construct the mapping relations between different kinds of features and the visibility of each scene. The experiment results show that the deep learned features exhibit better performance in both feature analysis and model training, especially NIMA achieves the best fitting performance with fewer feature dimensions.

A Sequential Machine Learning-cum-Attention Mechanism for Effective Segmentation of Brain Tumor.

Magnetic resonance imaging is the most generally utilized imaging methodology that permits radiologists to look inside the cerebrum using radio waves and magnets for tumor identification. However, it is tedious and complex to identify the tumorous and nontumorous regions due to the complexity in the tumorous region. Therefore, reliable and automatic segmentation and prediction are necessary for the segmentation of brain tumors. This paper proposes a reliable and efficient neural network variant, i.e., an attention-based convolutional neural network for brain tumor segmentation. Specifically, an encoder part of the UNET is a pre-trained VGG19 network followed by the adjacent decoder parts with an attention gate for segmentation noise induction and a denoising mechanism for avoiding overfitting. The dataset we are using for segmentation is BRATS’20, which comprises four different MRI modalities and one target mask file. The abovementioned algorithm resulted in a dice similarity coefficient of 0.83, 0.86, and 0.90 for enhancing, core, and whole tumors, respectively.

Segmentation of Echocardiography Based on Deep Learning Model

In order to achieve the classification of mitral regurgitation, a deep learning network VDS-UNET was designed to automatically segment the critical regions of echocardiography with three sections of apical two-chamber, apical three-chamber, and apical four-chamber. First, an expert-labeled dataset of 153 echocardiographic videos and 2183 images from 49 subjects was constructed. Then, the convolution layer in the VGG16 network was used to replace the contraction path in the original UNet network to extract image features, and depth supervision was added to the expansion path to achieve the segmentation of LA, LV, and MV. The results showed that the Dice coefficients of LA, LV, and MV were 0.935, 0.915, and 0.757, respectively. The proposed deep learning network can achieve simultaneous and accurate segmentation of LA, LV, and MV in multi-section echocardiography, laying a foundation for quantitative measurement of clinical parameters related to mitral regurgitation.

Deep Learning-Based Multi-Label Tissue Segmentation and Density Assessment from Mammograms

Objectives: Breast cancer is the most commonly diagnosed type of cancer among women and a common cause of cancer-related deaths. Early diagnosis and treatment of breast cancer is critical in disease prognosis. Breast density is known to have a correlation with breast cancer. In recent years, there has been an increasing interest in the investigation of computer-aided methods for early diagnosis of breast cancer. In this study, a new fully-automated deep learning-based cascaded model was proposed for breast density assessment. In the first stage, the segmentation of adipose, fibroglandular, and pectoral muscle tissues from the digitized film mammograms of the Digital Database for Screening Mammography (DDSM) was investigated using various types of U-nets. Features extracted from the breast tissue segmentation predictions were then used to assess breast density in the second stage. Material and methods: 66 and 296 mediolateral oblique mammograms were selected from DDSM dataset for segmentation and breast density assessment systems, respectively. Different U-nets with varying number of layers and filters were implemented and the model having the highest performance was determined. U-net performance was investigated using categorical cross-entropy, Dice, Tversky, Focal Tversky, and logarithmic cosine-hyperbolic Dice loss functions. The performances of U-nets having different types of connections were investigated. The performances of U-nets having pre-trained weights from VGG16, VGG19, and ResNet50 networks in the encoding path were also investigated. Segmentation results were improved by using an image processing pipeline based on morphological operators. Segmentation performance was presented in terms of accuracy, balanced accuracy, intersection over union, and Dice's similarity coefficient (DSC) metrics. The segmentation system predictions were then used to estimate mammographic density using a machine learning pipeline by extracting features related to the fibroglandular tissue percentage. Results: Using ResNet50-U-net on the test data, average DSC scores of 82.71%, 73.39%, and 95.30% were obtained for adipose, fibroglandular, and pectoral muscle tissue segmentation, respectively. The mammogram segmentation results are 3%-12% better than the current state-of-the-art DSC in the literature when considering all of the foreground tissues concurrently. A breast density classification accuracy of 76.01% was achieved on a separate mammogram dataset, which is comparable to the recent studies in the literature. Conclusion: The proposed system can be used for automatic segmentation of mammogram into adipose, fibroglandular, and pectoral muscle tissues. The segmentation model enables the estimation of the fibroglandular-adipose tissue interface, which is recently found to be an important region for breast cancer investigations. The proposed fully-automatic breast density assessment system has a comparable performance to the ones in the literature.

Scene classification for remote sensing images with self‐attention augmented CNN

Remote sensing scene classification aims to automatically assign a specific semantic label to each image. It is challenging to classify remote sensing scene images due to the images' diversity and rich spatial information. Recently, convolutional neural networks have been widely used to overcome these difficulties, such as the famous Visual Geometry Group (VGG) network. However, the VGG network with local receptive fields cannot model the global information of remote sensing images well. It also needs a large number of parameters and floating point operations to achieve satisfactory accuracy. To overcome these challenges, we introduce the self-attention mechanism to the VGG network. Specifically, we replace the last four convolutional layers in the VGG-19 network with two cascaded self-attention blocks, each consisting of two multi-head self-attention (MHSA) layers with the residual network structure. The new structure can simultaneously explore the local and global information from remote sensing scenes. Such improvements not only reduce model parameters but also improve the classification performance. The effectiveness of the proposed method is validated through experiments on four public data sets, i.e., NaSC-TG2, WHU-RS19, AID and EuroSAT.

Distributed detection of internal cavities in concrete-filled steel tube arch bridge elements

Internal cavities in structural elements, such as in concrete-filled steel tubes (CFSTs) of arch ribs reduce confinement, bearing capacity, and the durability of the arch bridges. Formation of internal cavities may be either materials related, which affects proper consolidation, or construction related, especially in terms of proper timing and delivery of concrete slurry to horizontal and inclined structural members. Nondestructive test methods such as ultrasonics have been employed for detection of internal cavities. Despite their success, point-by-point inspection of structural members becomes time-consuming and inefficient, especially when entire structures need to be inspected. Infrared thermography (IR) provides a more reasonable global approach for detection of anomalies. The accuracy and resolution of IR depend on the local ambient conditions affecting convection and external thermal flux prior to detection by IR. The research described herein pertains to the development of a hybrid thermographic approach by using Brillouin fiber optic sensors (BFOSs) for direct detection of thermal convection at the surfaces of steel tubes. Accurate quantitative detection of internal cavity locations and their dimensions required development of a specific machine learning-based approach through which the distributed thermographic data was converted to a series of grayscale images. A Faster Region-based Convolutional Neural Network (Faster R-CNN) based on the improved VGG-16 (IVGG-16) network architecture was utilized for this purpose. The experimental work involved design and fabrication of a distributed temperature sensing sheet (DTSS) for simultaneous transmission of thermal energy into CFST specimens and acquisition of distributed thermographic data by the BFOS. By using the proposed approach, it was possible to detect the embedded internal cavities during the experiments.

Convolutional Neural Network with Attention Module for Identification of Tunnel Seepage

As tunnel construction proceeds ever more rapidly, the efficiency of seepage detection by engineers with expert knowledge is facing unprecedented challenges. Moreover, it suffers from strong subjectivity. In recent years, deep learning, as an algorithm of machine learning, has achieved state-of-the-art performance in pattern recognition. In this paper, we address such a problem by building convolutional neural networks that operate on conventional graphics processing units. Within the project, the data is obtained by an infrared thermal imager since there exist different characteristics of temperature between the area of seepage and non-seepage. Considering the difficulty of collecting many images, generative adversarial nets and other data augmentation skills are applicable to enlarge data sets. We design several novel architectures where the attention mechanism is plugged into various traditional models, considered as VGG16 network with Attention Module and RestNet34 with Attention Module, and the overall identification accuracy achieved is more than 97%. The codes of this project can be found at https://github.com/Scotter-Qian/cnn .

Unsupervised despeckling of optical coherence tomography images by combining cross-scale CNN with an intra-patch and inter-patch based transformer.

Optical coherence tomography (OCT) has found wide application to the diagnosis of ophthalmic diseases, but the quality of OCT images is degraded by speckle noise. The convolutional neural network (CNN) based methods have attracted much attention in OCT image despeckling. However, these methods generally need noisy-clean image pairs for training and they are difficult to capture the global context information effectively. To address these issues, we have proposed a novel unsupervised despeckling method. This method uses the cross-scale CNN to extract the local features and uses the intra-patch and inter-patch based transformer to extract and merge the local and global feature information. Based on these extracted features, a reconstruction network is used to produce the final denoised result. The proposed network is trained using a hybrid unsupervised loss function, which is defined by the loss produced from Nerighbor2Neighbor, the structural similarity between the despeckled results of the probabilistic non-local means method and our method as well as the mean squared error between their features extracted by the VGG network. Experiments on two clinical OCT image datasets show that our method performs better than several popular despeckling algorithms in terms of visual evaluation and quantitative indexes.

A 3D Fluorescence Classification and Component Prediction Method Based on VGG Convolutional Neural Network and PARAFAC Analysis Method

Three-dimensional fluorescence is currently studied by methods such as parallel factor analysis (PARAFAC), fluorescence regional integration (FRI), and principal component analysis (PCA). There are also many studies combining convolutional neural networks at present, but there is no one method recognized as the most effective among the methods combining convolutional neural networks and 3D fluorescence analysis. Based on this, we took some samples from the actual environment for measuring 3D fluorescence data and obtained a batch of public datasets from the internet species. Firstly, we preprocessed the data (including two steps of PARAFAC analysis and CNN dataset generation), and then we proposed a 3D fluorescence classification method and a components fitting method based on VGG16 and VGG11 convolutional neural networks. The VGG16 network is used for the classification of 3D fluorescence data with a training accuracy of 99.6% (as same as the PCA + SVM method (99.6%)). Among the component maps fitting networks, we comprehensively compared the improved LeNet network, the improved AlexNet network, and the improved VGG11 network, and finally selected the improved VGG11 network as the component maps fitting network. In the improved VGG11 network training, we used the MSE loss function and cosine similarity to judge the merit of the model, and the MSE loss of the network training reached 4.6 × 10−4 (characterizing the variability of the training results and the actual results), and we used the cosine similarity as the accuracy criterion, and the cosine similarity of the training results reached 0.99 (comparison of the training results and the actual results). The network performance is excellent. The experiments demonstrate that the convolutional neural network has a great application in 3D fluorescence analysis.

An Improved Neural Network Model Based on Inception-v3 for Oracle Bone Inscription Character Recognition

Oracle bone inscription is the ancestor of modern Chinese characters. Character recognition is an essential part of the research of oracle bone inscription. In this paper, we propose an improved neural network model based on Inception-v3 for oracle bone inscription character recognition. We replace the original convolution block and add the Contextual Transformer block and the Convolutional Block Attention Module. We conduct character recognition experiments with the improved model on two oracle bone inscription character image datasets, HWOBC and OBC306, and the results indicate that the improved model can still achieve excellent results in the cases of blurred, occluded, and mutilated characters. We also select AlexNet, VGG-19, and Inception-v3 neural network models for the same experiments, and the comparison result shows that the proposed model outperforms other models in three evaluation indicators, namely, Top-1 Accuracy, Top-3 Accuracy, and Top-5 Accuracy, which indicate the correctness and excellence of our proposed model.

An Automatic Detection Method for COVID-19 in CT Images

There are many new cases of new coronary pneumonia (named COVID-19) every day around the world. In such a severe situation, effective detection of COVID-19 has become extremely important. Studies have shown that chest CT images can be used for COVID-19 detection because they can show bilateral changes in the lungs of people infected with COVID-19. It is not difficult for experienced radiologists to make preliminary judgments based on CT images. However, with the emergence of a large number of suspected cases, the explosive demand has overwhelmed doctors. Therefore, the automatic diagnosis of COVID-19 CT images is of great significance to realize early diagnosis, early isolation, and early treatment. In this paper, we study an automatic identification method that uses transfer learning technology, which transfers the pre-trained VGG16 network for feature extraction, and combines VAE augmentation data to reduce over-fitting, and finally uses integrated technology to achieve better detection results. Our experiments have shown that the accuracy of our method for identifying whether a patient’s CT image is positive or negative for COVID-19 is 91%, the precision is 88%, the recall rate is 94%, and the F1-score is 91%. Compared with other the state of art methods, the method proposed in this article can provide more efficient classification and identification of COVID-19.

Infrared and Visible Image Fusion in a Multilevel Low-Rank Decomposition Framework Based on Guided Filtering and Feature Extraction ∗

A novel infrared and visible image fusion method in a multilevel low-rank decomposition framework based on guided filtering and feature extraction is proposed to address the lack of edge information and blurred details in fused images. Based on multilevel low-rank decomposition, the fusion strategy of base part and detail contents has been improved. Firstly, the source infrared and visible images are decomposed to the base part coefficients and n -level detail content coefficients by multilevel low-rank decomposition. Secondly, the base part coefficients are learned by the VGG-19 network to get the weight map, and then, the improved weight map is obtained by guided filtering, and the coefficients of the base part are fused to acquire the fused base part coefficients. The n -level detail content coefficients are fused using the rule of dynamic level measurement with maximum value and then reconstructed to obtain the final fused detail content coefficients. Finally, the fused base part and detailed content information are superimposed to get the final fusion result. The results show that the fusion algorithm can effectively preserve the edge and detail features of the source image. Compared with other state-of-the-art fusion methods, the proposed method performs better in objective assessment and visual quality. The average value of evaluation metrics E N and M I have been improved by 0.5337 and 1.0673 on the six pair images.

BPI-MVQA: a bi-branch model for medical visual question answering

BackgroundVisual question answering in medical domain (VQA-Med) exhibits great potential for enhancing confidence in diagnosing diseases and helping patients better understand their medical conditions. One of the challenges in VQA-Med is how to better understand and combine the semantic features of medical images (e.g., X-rays, Magnetic Resonance Imaging(MRI)) and answer the corresponding questions accurately in unlabeled medical datasets.MethodWe propose a novel Bi-branched model based on Parallel networks and Image retrieval for Medical Visual Question Answering (BPI-MVQA). The first branch of BPI-MVQA is a transformer structure based on a parallel network to achieve complementary advantages in image sequence feature and spatial feature extraction, and multi-modal features are implicitly fused by using the multi-head self-attention mechanism. The second branch is retrieving the similarity of image features generated by the VGG16 network to obtain similar text descriptions as labels.ResultThe BPI-MVQA model achieves state-of-the-art results on three VQA-Med datasets, and the main metric scores exceed the best results so far by 0.2%, 1.4%, and 1.1%.ConclusionThe evaluation results support the effectiveness of the BPI-MVQA model in VQA-Med. The design of the bi-branch structure helps the model answer different types of visual questions. The parallel network allows for multi-angle image feature extraction, a unique feature extraction method that helps the model better understand the semantic information of the image and achieve greater accuracy in the multi-classification of VQA-Med. In addition, image retrieval helps the model answer irregular, open-ended type questions from the perspective of understanding the information provided by images. The comparison of our method with state-of-the-art methods on three datasets also shows that our method can bring substantial improvement to the VQA-Med system.

A improved detection method for lung nodule based on multi‐scale 3Dconvolutional neural network

AbstractLow detection sensitivity and high false positives are two main challenges from traditional lung nodule detection model. In order to improve reliability of detection model, we proposed a multi‐scale three‐dimensional convolutional neural network lung nodule detection method which can simultaneously detect lung nodules and reduce false positives. First, a 3D UNet++‐like architecture with an encoding‐decoding structure is regarded as a feature extraction network and combines it with the region proposal network. The multi‐feature fusion method can fully learn features at different levels through the cross‐layer connection between the encoder and the decoder. Second, the layers of the 3D UNet++‐like architecture are connected by residual blocks, which enhance feature reuse and speeds up the convergence speed of the network. Finally, the three input sizes are input into the 3D neural network, and their classification results are merged, thus to find the final nodule determination result. Based on the LUNA16 dataset, the experiment result shows that average sensitivity is 87.3% and be increased 7.8% than the UNet++ network in our model. And if the number of candidate nodules is 48, the sensitivity is as high as 96.2%, it increases 8.1% than the VGG16 network. Obviously, that the improved model can significantly improve detection sensitivity and reduce false positives, which can provide a valuable theoretical reference for clinical medicine.

Lightweight dense-scale network (LDSNet) for corn leaf disease identification

The identification of corn leaf diseases in real scenarios faces important challenges, such as complex background interference, intra- and inter-class scale changes, and lightweight model deployment. To overcome these challenges, we propose a lightweight dense-scale network (LDSNet) for real-world corn leaf disease image identification. The main component of LDSNet is the improved dense dilated convolution (IDDC) block that has two key improvements relative to existing essential blocks. The first one is that it improves the adaptability to the scale change of corn leaf diseases through the dense connection of different dilation rate convolutions. The second improvement is that it replaces the concatenation connection with a new fusion method, namely, coordinated attention scale fusion, for enhanced extraction of corn leaf features in a complex background. In addition, we propose a new loss function to optimize the LDSNet network model. Experimental results show that the accuracy of the optimized model on the test data set reaches 95.4%, which is better than the accuracy of existing heavyweight networks, such as AlexNet, VGG16, VGG19, and ResNet50, and lightweight networks, such as DenseNet121, GoogleNet, MobileNet (V1, V2, and V3-large), ShuffleNetV2, and GhostNet. The number of parameters accounts for only 45.4% of the minimum number of parameters (ShuffleNetV2, 1.3M) in the compared model. To verify the practical application performance of the proposed network model, we apply the trained model to a mobile phone. A real-world test proves that our model has strong compatibility and high recognition performance.

ASTERI: image-based representation of EEG signals for motor imagery classification

PurposeElectroencephalography (EEG) signals are valuable in the monitoring and investigation of neurological diseases and in the control of brain-machine interfaces (BCI). However, these signals are noisy and are non-linear and non-stationary in nature. Signal analysis is an expensive task and can lead to misdiagnosis. Deep learning can be used to overcome these challenges. The most used deep architectures are based on convolutional neural networks (CNNs). Representing EEG signals as images can be useful to use deep architectures based on CNNs in solutions based on intelligent EEG-based systems.MethodsIn this work, we propose the ASTERI method, for representing EEG signals in two dimensions using the backprojection reconstruction method. To validate the proposal, experiments were performed with motor imagery EEG signals from the BCI Competition IV 2b motor imagery database. To extract attributes from the EEG signal windows, we used the ASTERI representation and investigated the pre-trained networks VGG16 and LeNet. To classify the final classification of imagined movements to the right and to the left, we used Random Forests.ResultsFor the original database, the validation with 100 trees achieved an average accuracy and kappa index of 88.97% and 0.78, respectively. In the case of augmented databases with synthetic instances, the average accuracies were 89.10% and 89.13% for increases of 50% and 150% for each class, respectively.ConclusionThe experimental results showed that the ASTERI method, together with hybrid deep architectures based on CNN and Random Forests, can generate competitive results in solving motor imagery problems. The representation of signals as images can also contribute to the exploration of new solutions of applied neuroscience problems based on deep architectures.

Research on Automatic Cargo Recognition in Smart City Environment

Smart city refers to the use of various information technologies to improve the lives of citizens. However, in terms of transportation and sales of goods, traditional methods require a lot of manpower and material resources, and cannot be automatically identified. In order to improve the efficiency and accuracy of product identification, product sorting is automated. It uses the powerful feature learning and expression capabilities of deep convolutional neural networks to automatically learn product features, thereby achieving high-precision image classification. Therefore, this paper first proposes an improved VGG network, combines transfer learning to establish a deep learning recognition model, and finally conducts multiple sets of experiments on the 131-category Fruit-360 dataset. The results show that when the Adam optimizer is used for iterative training for 30 rounds and the batch_size is 64, the accuracy of the algorithm proposed in this paper reaches 94.19% on the training set, 97.91% on the validation set, and 92.2% on the test set top1. The accuracy rate on the test set top5 is as high as 100%. Therefore, the method in this paper can solve the problems caused by traditional methods and provide useful help for smart cities.