Search

In the era of big data, with the continuous popularization of social media, various text data are growing in the network and in life. It is of great significance to analyze and manage text data using text classification technology. Text classification is a basic research field in the field of artificial intelligence natural language processing. Under the given criteria, it classifies text according to content. The application scenarios of text classification are very extensive, such as sentiment analysis, topic classification, relationship classification, etc. Deep learning is a method of representation learning based on data in machine learning, and it shows good classification effect in text data processing. Chinese text and English text have differences in form, sound, and image. Focusing on the uniqueness of Chinese text classification, this paper analyzes and expounds the deep learning methods used for Chinese text classification, and finally sorts out commonly used datasets for Chinese text classification. [ABSTRACT FROM AUTHOR]

In industrial production, defects are treated differently based on their severity, so it is necessary to classify defects. However, in actual production, the classification accuracy is often insufficient due to the presence of few easily confused defects, which requires conservative treatment of all defects in production practice, resulting in significant human and economic costs. To solve this problem, this paper proposes a method for interactive separating easily confused defects. This method separates few easily confused defects from other defects, ensuring that the classification results of the remaining majority of defects can be directly used. This method selects easily confused defects from the training data as one or more new defect categories, called virtual defects, so that the trained network can distinguish between virtual defects and other defects. This paper designs a visual interface to assist users in interactively selecting easily confused defects to construct virtual categories. CMOS defect data from actual industrial sites are adopted for effectiveness verification, and the results show that the proposed method can quickly classify few confusing defects and ensure that the classification accuracy of remaining defects meets the requirements of industrial applications. [ABSTRACT FROM AUTHOR]

With the development of artificial intelligence, digital images have been widely used in various fields. However, due to the appearance of image editing software, a large number of images have been tampered with maliciously, which seriously affects the authenticity of image content. Different from the general object detection, the study of image tampering detection needs to pay more attention to the tamper information of the image itself, which is often manifested in a weak form. Therefore, the detection model needs to focus on learning more abundant tamper features. This paper proposes a dual-stream Faster R-CNN model that combines gradient edge information and attention mechanism, and the model can realize detection and location of regions with different tampering types. One of the two streams is the color stream, which uses the attention mechanism to extract the surface features of the image, such as brightness contrast, visual difference of tampering with the boundary, etc. The second of the two streams is a gradient stream. A Gradient high-pass filter is used to enhance the anomaly edge features between the real area and the tampered area, making it easier for the model to find faint tampered traces in the tampered image. Finally, the features of color stream and gradient stream are fused by means of compact bilinear pooling. Due to the relatively small size of publicly available image tampering data sets, the Pascal VOC 2012 is used to create an image tampering detection data set which containing 10 010 images for model pre-training. The experimental results on COVER, Columbia, and CASIA data sets show that the model proposed in this paper improves the detection accuracy by 7.1% to 9.6% compared to the latest models, and exhibits higher robustness under JPEG compression and image blur attacks. [ABSTRACT FROM AUTHOR]

Copyright of Computer Engineering & Science / Jisuanji Gongcheng yu Kexue is the property of Computer Engineering & Science and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Copyright of Computer Engineering & Science / Jisuanji Gongcheng yu Kexue is the property of Computer Engineering & Science and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

To meet the fast iteration speed and high computing power requirements of deep learning models, mainstream hardware vendors are increasingly inclined towards heterogeneous systems consisting of general-purpose processors and AI-specific accelerator cores. However, AI-specific accelerator ores only support certain core operators and do not have general programming capabilities. Therefore, how to efficiently deploy deep learning tasks on such heterogeneous architectures is worth further research. Based on the domestically developed FT-Matrix DSP+MatrixZone heterogeneous system platform, this paper designs and implements a deep learning programming framework, called KaiSa. KaiSa analyzes the input parameters of the deep learning model, identifies the operator type, and assigns it to the corresponding computing core. For complex operators, KaiSa automatically completes the optimal search for the block size based on a performance model, improving the performance of dual-core parallel computing. At the same time, KaiSa shields all low-level hardware details to provide users with a friendly programming environment for efficient program development. Experimental results show that KaiSa can achieve performance improvements of up to 39.0%. [ABSTRACT FROM AUTHOR]

Network intrusion detection systems are crucial for maintaining network security, and there is currently limited research on intrusion detection scenarios with only a few abnormal markers of network data. This paper designs a weakly-supervised learning intrusion detection model, called WIDSAPL, based on the anomaly retention of data. The detection model consists of four parts: data transformation layer, representation learning layer, transformation classification layer, and anomaly discrimination layer. By using a set of learnable encoders to map samples to different regions and compress them into a hypersphere, the label information of abnormal samples is used to learn the classification boundaries of normal and abnormal samples, and the abnormal score of the samples is obtained. Testing the WIDS-APL system on four datasets demonstrates the effectiveness and robustness of the system, with improvements in the AUC-ROC values of 4.80%, 5.96%, 1.58%, and 1.73% respectively compared to other mainstream methods. Furthermore, there are enhancements of 15.03%, 2.95%, 4.71%, and 9.23% in AUC-PR performance. [ABSTRACT FROM AUTHOR]

Copyright of Computer Engineering & Science / Jisuanji Gongcheng yu Kexue is the property of Computer Engineering & Science and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The large amount of medical information carried in the electronic medical record can help doctors better understand the situation of patients and assist doctors in clinical diagnosis. As the two core tasks of Chinese electronic medical record (EMR) information extraction, named entity recognition and entity relationship extraction have become the main research directions. Its main goal is to identify the medical entities in the EMR text and extract the medical relationships between the entities. This paper systematically expounds the research status of Chinese electronic medical record, points out the important role of named entity recognition and entity relationship extraction in Chinese electronic medical record information extraction, then introduces the latest research results of named entity recognition and relationship extraction algorithm for Chinese electronic medical record information extraction, and analyzes the advantages and disadvantages of each model in each stage. In addition, the current problems of Chinese EMR are discussed, and the future research trend is prospected. [ABSTRACT FROM AUTHOR]

The development of medical imaging technology has generated a massive amount of medical image data, which reflects the internal structural features of the human body. Medical image segmentation technology can improve the efficiency of medical diagnosis, making it an important assistive tool for modern medical diagnosis. However, noise or artifacts that are inevitably present in the imaging process bring great challenges to the segmentation work. In existing segmentation models, single-frame medical image semantic segmentation models do not consider the relationship between image frames, while video semantic segmentation models utilize temporal information but have some limitations in edge extraction. To address these issues, this paper proposes a U-Net-based temporal semantic segmentation model modified by optical flow. This model can extract optical flow information between consecutive frames and perform feature extraction and weight allocation on the current frame and optical flow for correction. Experiments show that the model obtains optimal results on three evaluation metrics, namely Dice similarity, pixel accuracy and cross-merge ratio, on different types of datasets, namely Drosophila electron micrographs, combined healthy abdominal organ segmentation and coronary angiogram, which validate the effectiveness and generalization of the proposed model. [ABSTRACT FROM AUTHOR]

Copyright of Computer Engineering & Science / Jisuanji Gongcheng yu Kexue is the property of Computer Engineering & Science and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Surface crack identification with visual images is a kind of non-contact detection solution, which is not limited by the material of the tested object and is easy to achieve online automation. Therefore, it has the advantages of fast speed, low cost and high precision. Firstly, the public data sets of typical pavement crack are comprehensively collected, and the characteristics of sample images and the random variable factors are summarized. Subsequently, the advantages and disadvantages of hand-crafted feature engineering, machine learning, and deep learning crack identification methods are compared. Finally, from the aspects of network architecture, testing performance and predicting effectiveness, this paper reviews the development and progress of typical deep learning algorithms such as self-built CNN, transfer learning (TL) and encoder-decoder (ED) that can be easily trained and deployed. We can see the obvious improvement of performance and effect because of algorithm optimization and computing power enhancement. The results show that real-time patch-level and fast pixel-level crack detection can be realized on low computing power GPU platform. [ABSTRACT FROM AUTHOR]

Precipitation nowcasting refers to the high-resolution prediction of precipitation in the short term, which is an important but difficult task. In the context of deep learning, it is viewed as a radar echo map-based spatiotemporal sequence prediction problem. Precipitation prediction is a complex self-supervised task. Since the motion always changes significantly in both spatial and temporal dimensions, it is difficult for ordinary models to cope with complex nonlinear spatiotemporal transformations, resulting in blurred predictions. Therefore, how to further improve the model prediction performance and reduce ambiguity is a key focus of research in this field. Currently, the research on precipitation nowcasting is still in the early stage, and there is a lack of systematic classification and discussion about the existing research work. Therefore, it is necessary to conduct a comprehensive investigation in this field. This paper comprehensively summarizes and analyzes the relevant knowledge in the field of precipitation nowcasting from different dimensions, and gives future research directions. The specific contents are as follows: (1) The significance of precipitation nowcasting, and the advantages and disadvantages of traditional forecasting models are clarified. (2) The mathematical definition of the nowcasting prob-lem is given. (3) Common predictive models are comprehensively summarized, analyzed. (4) Several open source radar datasets in different countries and regions are introduced, and download links are given. (5) The metrics used for prediction quality assessment are briefly introduced. (6) The different loss functions used in different models is discussed. (7) The research direction of precipitation nowcasting in the future is pointed out. [ABSTRACT FROM AUTHOR]

Aiming at the erroneous literacy caused by irregular interference and text adhesion in text images, this paper proposes a text image restoration model based on partial convolution operations. This paper studies and analyzes the text image characteristics of several common fonts, establishes a text image database, and integrates it with the interference mask database, then evaluates the repair effect of the model, and conducts classification tests on different levels of repair. Experiments have proved that the text model predicts the missing parts based on the existing parts of the current text under the premise of ensuring that the original text information is not lost. The peak signal-to-noise ratio is up to 32. 46 dB, the structural similarity is up to 0. 954, and the best loss value is up to 0. 015. The text recognition rate after repair is 27. 85% higher than that before repair. After repairing the defective pictures of four ancient scripts, including official script, seal script, oracle bone inscriptions, and running script, the peak signal-to-noise ratio reached 30. 46 dB, and the structural similarity reached 0. 964. [ABSTRACT FROM AUTHOR]

The curvilinear motion of solar photosphere bright points, which is approximate rotation, is of great significance for studying how the energy from solar convection zone is transmitted to the corona. The existing algorithms only detect the global curvilinear motion of photosphere bright points. This paper proposes a multi-scale and multi-modal deep learning method to detect the global and local curvilinear motion of photosphere bright points. This mothod constructs a multi-scale network model based on the bidirectional long short-term memory network (Bi-LSTM) to extract multi-scale time sequence features of the trajectories of photosphere bright points. EfficientNet-B0 is adopted to extract the spatial features of the trajectories. The temporal features and spatial features are fused into multi-modal features to detect various curvilinear motions of photosphere bright points. The experiment results show that the accuracy of this method is 85.08%, which is 6.12% higher than that of the single-scale method and 3.1% higher than that of the multi-scale and single-mode method. This method can also be applied to the motion type detection requirements in other fields. [ABSTRACT FROM AUTHOR]

Alzheimers disease (AD) is an irreversible neuro degenerative brain disease and the most common dementia in the elderly. Manual classification of Alzheimer's magnetic resonance image (MRI) has problems delay classification and time-consuming classification. As the aging population becomes more and more serious accurately and quickly classify patients with AD has important research significance. This paper combines convolutional neural network (CNN) technology with MRI technology and designs a 3D-ResNet model for AD classification which achieves 98. 39% accuracy 96.74% sensitivity and 99. 99% specificity on the validation set and achieves 97. 43% accuracy 94. 92% sensitivity and 99. 99% specificity on the test set. The classification time of each patient is 0. 23 s. In addition，for the problem that the pathogenesis of AD is not yet clear，this paper uses Class Activation Mapping (CAM) technology to visualize the AD-related brain regions. [ABSTRACT FROM AUTHOR]

Copyright of Computer Engineering & Science / Jisuanji Gongcheng yu Kexue is the property of Computer Engineering & Science and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Generating images from text is a challenge task in machine learning community. Although significant success has been achieved so far, problems such as unstable network training and disappearing gradients still exist. In response to the above shortcomings, based on the stacked generative confrontation network model (StackGAN), this paper proposes a text-to-image generation method that combines spectral normalization and perceptual loss function. Firstly, the network model applies spectral normalization to the discriminator, restricts the gradient of each layer of the network to a fixed range, slows down the convergence speed of the discriminator, and hence improves the stability of network training. Secondly, the perceptual loss function is added to the generator network to enhance the consistency between the text content and the generated image. The network model uses Inception scores to evaluate the quality of the generated images. The experimental results show that, compared with the original StackGAN, the network model has better stability and generates clearer images. [ABSTRACT FROM AUTHOR]

Crowd counting and crowd density estimation are the important branch in crowd analysis， and are also the important information that surveillance always concerns. Although some important progress has been made in this field in recent decades, there are still some challenging problems. This paper reviews the current researches and development trends of the crowd counting and crowd density estimation methods based on computer vision. Firstly，the development background and application direction of the crowd counting and crowd density estimation technology are introduced. Secondly, the important methods proposed in recent years are summarized, which can be divided into two types from the machine learning point of view: the shallow learning based methods and the deep learning based methods. In the other hand, they can be divided into two types from learning model standpoint： the direct method (i. e.， the detection based method) and the indirect method (i. e.，the pixel-based, texture-based and corner point based methods). This paper introduces the shallow learning based methods in the last twenty years in detail, and makes a brief summary of the deep learning based methods in recent years. Then, a brief introduction is made on the performance evaluation techniques of crowd counting and crowd density estimation methods, and several data sets are provided for testing and evaluating these methods. Finally, the technical challenges in the field are summarized and the future research directions are prospected. [ABSTRACT FROM AUTHOR]

The deployment environment of distributed photovoltaic power plants is relatively complicated, and many kinds of faults inevitably occur during the actual operation. In order to solve the above problem, this paper proposes a fault diagnosis model of distributed photovoltaic power stations based on deep residual network. It analyzes and processes the sequence data of equipment operation, and achieves rapid and accurate judgment of fault categories. This model applies a one-dimensional convolution kernel to perceive the characteristics of time series data. Then, it uses a multi-level convolution structure to increase the diagnostic ability. Finally, the residual network is utilized to solve the problem of gradient disappearance caused by the increase of model depth, and accelerate the training of the deep model. The experimental results based on the power station test data show that the residual network model achieves higher fault diagnosis accuracy than several st ate-of-the-art intelligen t models. The application of this model can not only greatly reduce the investment in fault inspection of photovoltaic power plants, but also improve the efficiency of fault diagnosis of photovoltaic power plants. [ABSTRACT FROM AUTHOR]

Sketch retrieval (SBIR) is an extension of content-based image retrieval (CBIR), which is a flexible and convenient way to retrieve target images. How to minimize the difference between the sketch domain and the image domain is crucial to SBIR. The traditional methods extract the manual fea-tu res to achieve the approxima te conversion bet ween the sketch field and the image field, so as to reduce the domain difference. However, these methods cannot effectively fit the content of the two domains, resulting in low retrieval accuracy. Deep learning methods break through the limitations of traditional methods, which extract high-dimensional features from a large amount of data and have been proved to effectively solve the cross-domain modeling problems. This paper focuses on deep learning-based sketch retrieval methods, and covers several aspects such as the deep feature extraction model, public dataset, coarse-grained and fine-grained retrieval based on deep learning, deep hashing technology, category generalization, etc. Related works are reviewed and commented on. Then, a comparative experiment is conducted. For one hand, three existing public SBIR datasets such as Sketchy, TU-Berlin and QuickDraw are used for suitability evaluation. For the other hand, three latest SBIR deep learning models such as GRLZS model, SEM-PCYC model and SAKE model are selected for performance analysis and comparison. Finally, current challenges and future research trends of SBIR are summarized. [ABSTRACT FROM AUTHOR]

Behavior det ection is a research hot spot in the field of video understanding and computer vision, which attracts the attention of scholars at home and abroad. It has been widely used in many fields such as intelligent surveillance and human-computer interaction. With the development of technology, deep learning has made a great breakthrough in image classification. The application of the recognition methods based on deep learning to human behavior detection has become a hotspot. Therefore, the paper firstly introduces several datasets commonly used in behavior detection, and the research status of deep learning in the field of behavior detection in recent years. Then, the basic process of behavior detection methods and recognition methods based on deep learning are analyzed. Finally, the future development trend and possible shortcomings are analyzed from the aspects of method performance and application prospect. [ABSTRACT FROM AUTHOR]

Ocean mesoscale eddy is an important ocean mesoscale phenomenon that plays an important role in ocean circulation, mat erial and energy transport, and has an important impact on the safety of ship navigation and hydroacoustic communication. Efficient and accurate detection and identification of ocean mesoscale eddies are of great research value for both physical ocean cognition and ocean exploitation. Traditional eddy detection and identification methods rely on a single threshold designed by experts experiences. With the rise of deep learning, the current machine learning methods show certain advantages in the accuracy and automation of eddy detection and identification. This paper summarizes and comparatively analyzes the existing machine learning-based detection and identification methods to provide a systematic knowledge and reference basis for the development of ocean mesoscale eddy detection and identification research. [ABSTRACT FROM AUTHOR]

Semantic segmentation algorithms can classify images at the pixel level, and are widely used in fields such as unmanned driving, medical image processing, and industrial automation, and have important research value. The research of semantic segmentation algorithms focuses on three aspects: improving the accuracy of segmentation, reducing the amount of parameter s and increasing the speed of inference. The lightweight semantic segmentation algorithm ENet uses a multi-layer convolutional codec and a large number of dilated convolutions to avoid excessive downsampling and use of spatial information. Although it retains some spatial information integrity and large receptive field, the codec is bloat-ed, the transmission of spatial information is poor, and the sensory field overflows and causes gr id effect. Aiming at the above problems, this paper tailors the EN et algorithm structure, uses the attention mechanism and the pyramid dilated convolution to design spatial information transmission module, optimizes the algorithm structure, improve s the algorithm receptive field, and completely transmits the spatial information transmission. The experimental results on public datasets Cityscapes and BDD100K show that the new module can improve the performance of the original algorithm with a smaller amount o f parameters and calculations, which proves the redundancy of t he original algorithm and the effectiveness of the designed module. [ABSTRACT FROM AUTHOR]

The Deep Neural Network (DNN) model is an important branch of the Artificial Neural Network (ANN) model and the foundation of deep learning. In recent years, due to the improvement of computer computing power and the development of high-performance computing technology, it has become possible to increase the DNN network depth and the model complexity to improve its feature extraction and data fitting capabilities. As a result, DNN has shown advantages in natural language processing, autonomous driving, face recognition and other Issues. However, big data and complex models have greatly increased the training cost of deep neural networks. Therefore, accelerating the training process has become a key task. Its technical scope covers many aspects from the design of the underlying current to the design of distributed algorithms. The peak speed of the domestic Taanhe aimed at one quintillion of times, and the huge computing power provides a potential opportunity for DNN training. Based on the characteristics of the ARM architecture of the Taanhe prototype, using the PyTorch framework and MPI technology, this paper conducts a uniquely designed CNN training for a sing FT 2000 3 computing node, a single MT2000 3 computing node, and the multi-node cluster expanded through them. The performance of the above-mentioned processors in neural network distributed train-ing has been optimized and evaluated, which provides experimental data and theoretical basis for further improving the performance of the Tianhe-3 prototype system in neural network distributed training. [ABSTRACT FROM AUTHOR]

Sperm head shape is an important indicator in the analysis of sperm morphology, which is very important for diagnosing male infertility. Therefore, it is very important to segment the sperm head accurately and efficiently. Based on this, this paper builds a new encoder-decoder segmentation network that combines stacked residual block and residual hybrid dilated convolution. Firstly, we build a dataset for segmenting the head of human sperm, which contains 1207 images, and then use it to train and test our network. The proposed network is able to achieve excellent segmentation results in low-quality images that are unstained and contain multiple sperms, and obtain a Dice coefficient of 96. 06% on the validation set. The experimental results show that the stacked residual module and the residual hybrid dilated convolution module significantly improve the segmentation performance. In addition, the proposed network processes the images that show the original true state of the sperm, and the accurate segmentation results are very helpful for the doctor's clinical diagnosis. [ABSTRACT FROM AUTHOR]

Single image super-resolution (SISR) refers to the recovery of a high-resolution image from a single low-resolution image. With deep learning used in the field of image super-resolution, deep networks can independently learn the mapping relationship between low-resolution and high-resolution training images, showing better reconstruction performance than the traditional methods. Therefore, deep learning has become dominant in super-resolution. This paper focuses on the exploration of the existing deep network model of super-resolution in terms of reconstruction mode, network structure, and loss function. By comparing the similarities and differences between different models, the advantages and disadvantages of different model building methods and the applicable application scenarios are analyzed. Meanwhile, the reconstruction results of different network models on the benchmark test datasets are compared and the potential directions are concluded. [ABSTRACT FROM AUTHOR]

Deep learning has achieved remarkable results in the field of facial recognition. However, when dealing with facial images under complex conditions such as occlusion, illumination and improper angles, predicting a large number of facial feature points is still a challenging problem. In order to solve the localization problem of multiple facial feature points under complex conditions, this paper designs a network structure based on C-Canny algorithm and improved single neural network. The traditional Canny algorithm is applied to the face region localization stage, so that the neural network can quickly reposition the face region to improve the accuracy of model recognition. Experimental results show that, compared with some existing traditional algorithms and neural networks, the neural network structure reduces the value of loss function by 12.2% on average on the 300-w and 300-vw datasets. [ABSTRACT FROM AUTHOR]

Improving the generalization ability and reducing the over-fitting risk is the research focus of deep convolutional neural networks. Occlusion is one of the critical factors affecting the generalization ability of convolutional neural networks. It is usually hoped that the models after complex training can have a good generalization for occlusion images. In order to reduce the over-fitting risk and improve the robustness of the model to random occlusion image recognition, this paper proposes an activation feature processing algorithm. During the training process, the input image is occluded by processing the maximum activation feature map of a convolutional layer, then the occluded new image is used as a new input to the network to go on training the model. The experimental results show that the proposed algorithm can improve the classification performance of multiple convolutional neural network models on different datasets and the trained models have excellent robustness to the identification of random occlusion images. [ABSTRACT FROM AUTHOR]

The traditional Unmanned Aerial Vehicle (UAV) human-machine interaction requires specialized equipment and professional training, and convenient and innovative ways of interaction are often more popular. In this paper, with ordinary cameras, we study the UAV gesture control system based on computer vision and deep learning. The system first uses the fast tracking algorithm to extract the operator's region in the video sequence, greatly reducing the pressure of subsequent video processing while removing the influence of complex background and camera drift. Secondly, according to the time information of the actions, the optical flow features are encoded in different colors and superimposed on a picture, and the video is converted into a color texture map that contains both temporal features and spatial features. Finally, colored texture images are well learned and classified by a deep Convolutional Neural Network (CNN) and UAV controlling commands are generated according to the classified results. The proposed system estimates actions within 1.6s every 0.4s and uses CNN to classify pictures so as to achieve real-time human-computer interaction. The system has a recognition accuracy of over 93% within 60 meters. In indoor and outdoor environments, the operator can conveniently control the UAV by imitating command actions. [ABSTRACT FROM AUTHOR]

Aiming at the problem that the current clothing classification algorithm has general accuracy in solving the multi-category clothing classification, this paper proposes an optimized clothing classification algorithm based on residual convolutional neural network, and uses the following three optimization methods in the network： 1) The orders of batch normalized layer (BN)，activation function (Relu) and convolution layer in the network are adjusted； 2) A parallel pooling structure of "pool layer + convoluted layer" is adopted； 3) The full connection layer is replaced by the global mean pooling layer. Experiments on the multi-category large-scale clothing data set (DeepFashion) provided by the multimedia laboratory，the Chinese university of Hong Kong and the standard data set CIFAR-10 show that the proposed network model is superior to VGGNet and AlexNet in image processing speed and classification accuracy, and obtains the best classification accuracy on DeepFashion data set so far. The network can also be better applied to target detection and image segmentation. [ABSTRACT FROM AUTHOR]