Search

Convolutional neural networks have achieved great success in the field of image analysis. With the development of deep learning, deep learning models are becoming more and more complex, and he amount of deep learning calculations is increasing rapidly. The sparse algorithm can effectively reduce the amount of deep learning calculations without reducing the accuracy. This paper uses three different pruning strategies under the ResNet18 model and GoogleNet model to reduce the calculation amount of the deep learning model. The results show that the global unstructured pruning strategy has a sparsity of 94% and 90% without reducing the accuracy respectively, the level unstructured pruning strategy has an average sparsity of 83% and 56% without basically reducing the accuracy respectively, and the level structured strategy has an average sparsity of 34% and 22% without basically reducing the accuracy respectively. Under the three pruning strategies, the delay and power consumption results obtained by running the sparse deep learning model in the Eyeriss deep learning accelerator shows that, compared with the unpruned strategy, under the ResNet model, the global unstructured pruning strategy has a 66.0% reduction in latency and a 60.7% reduction in power consumption, the level unstructured pruning strategy has a 66.0% reduction in delay and a 80.6% reduction in power consumption, and the level structured pruning strategy has a 65.6% reduction in latency and a 33.5% reduction in power consumption. Under the GoogleNet model, the global unstructured pruning strategy has a 4.5% reduction in latency and a 63.2% reduction in power consumption, the level unstructured pruning strategy has a 73.6% reduction in delay and a 55.0% reduction in power consumption, and the level structured pruning strategy has a 26.8% reduction in latency and a 5.8% reduction in power consumption. Therefore, this paper concludes that the global unstructured pruning strategy can greatly reduce he delay and energy consumption without reducing the accuracy. Under the level unstructured pruning strategy, the delay and energy consumption can be greatly reduced under the premise of slightly reducing the accuracy. [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 sequence recommendation algorithm based on self-attention mechanism utilizes users' interactive sequences to model their dynamic preferences and predict their future behaviors. However, directly inputting the interactive sequences into the self-attention layer will limit the effective utilization of local association information in the sequences. In add-on, most of the existing recommendation algorithms use the dot product of the representation of the users' recent behaviors and the target items to calculate the item scores, which will weaken the impact of previous interactive items on the recommendation results. This paper proposes a date-aware sequential recommendation algorithm that fuses local information of sequences. It uses multiple vertical filters to fuse multiple local association information of each interactive item in the interactive sequence, and uses cross-attention mechanism to capture the relationships between all historical interactive items and target items. It also abandons the traditional position embedding method and replaces it with absolute time embedding based on the date of interaction occurrence. Experimental results on multiple public datasets show that the algorithm has certain improvement compared with the baseline algorithms in different evaluation indicators. [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.)

To accurately recognize and evaluate tennis actions, by combining computer vision with tennis related knowledge, this paper proposes a tennis action recognition and evaluation method based on PoseC3D. Firstly, a pose estimation model based on resnet-50 is used to detect human targets in tennis video and extract bone key points. Secondly, the PoseC3D model is trained through the video data set collected in the professional tennis court, so that it can classify the sub actions of tennis. Thirdly, the dynamic time warping algorithm is used to evaluate the classified actions. Finally, based on the collected video data set, a large number of experiments are carried out. The results show that the Top1 accuracy of the proposed tennis action recognition method based on PoseC3D can reach 90.8%. Compared with the methods based on graph convolution network, such as AGCN and ST-GCN, it has stronger generalization ability. Moreover, the proposed scoring algorithm based on dynamic time warping can give real-time and accurate evaluation scores for corresponding actions after action classification, reducing the work intensity of tennis teachers and effectively improving the quality of tennis teaching. [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]

At present deep learning has played an important role in various human activity recognition (HAR) tasks. However, the activity data has the particularity of time series and includes body movements. The existing convolutional neural network (CNN) will cause the data to be highly correlated when performing convolutional operations. As the network affects the next layer, the accuracy of network recognition is limited. In order to solve this phenomenon, this paper proposes an improved convolutional neural network with covariance matrix for HAR scenario. It builds a de-correlated network structurethrough matrixtransformationtoeliminatecorrelationproblems. When the network performance is poor, the network can replace the existing BN layer to normalize data. The verification experimentsarefinishedonfourHARpublicdatasets. The proposedneural network is compared withtraditionalCNN modelandBNlayermodel. The results show that the improved neural network is improved by 1% to 2% compared with the previous deep learning networks, which proves that the improved neural network is effective. Furthermore, the application is transplanted to the mobile terminal for real-time activity recognition. [ABSTRACT FROM AUTHOR]

Social media Key Opinion Leader (KOL) is very popular with advertisers because of their excellent business value. However, with the low entry threshold of the KOL industry and data fraud, advertisers are unable to find a KOL that matches their own brand quickly. Based on the above background, this paper studies the video released by KOL on social platforms, analyzes the dynamic theme of the barrage text in the video, and describes the change of the theme of the barrage over time. At the same time, a convolutional neural network model is used to perform sentiment analysis on the barrage text of the video containing the advertisement, and further analyze the audience's emotional polarity to the situation that the video released by KOL containing the advertisement. The experimental results show that the proposed KOL analysis method evaluates the commercial value of KOL more comprehensively and specifically, helping advertisers find a suitable KOL efficiently. [ABSTRACT FROM AUTHOR]

YOLOv3-tiny has the excellent target detection capability, but the computational power required by the model is still large, so it is difficult to be used in the embedded application field. This paper proposes a hardware acceleration method of YOLOv3-tiny and implements it on FPGA platform. Firstly, for the fixed-point design of the network, with data accuracy and resource consumption as design indicators, through the statistics of the data distribution in the model and the division of data types, different fixed-point strategies are determined. Secondly, for the parallel design of the network, through the analysis of the calculation characteristics of the convolutional neural network, with the methods of loop adjustment, loop block, loop expansion, and array splitting, a scalable common hardware computing unit is designed. Then, for the network pipeline design, the research is carried out from two aspects: the inter-layer and the intra-layer. Based on the direction of the inter-layer data flow and the division of tasks within the layer, a flexible pipeline computing architecture is designed. Lastly, on the XILINX XC7Z020CLG400-1 platform, experiments demonstrate that, compared with single-core ARM-A9 processor at 667MHz, the proposal achieves the calculation speed as high as 290. 56. [ABSTRACT FROM AUTHOR]

Abdominal artery segmentation is an essential task for the diagnosis of gas trie cancer lymph node metastasis and the judgment of hepatic artery variant type. In order to solve the problems of low segmentation accuracy and easy fracture of abdominal artery, this paper proposes an abdominal artery segmentation method based on improved convolutional neural network. A pre-training module (resnet34) with convolutional attention is employed in encoding part of convolutional network to avoid the disappearance of gradients and better obtain the feature information of the images. In order to expand the receptive field and gather multi-scale feature information, a new multi-scale feature fusion module is proposed. In addition, the learning of the edge structure information of arteries is very significant. Attention guide filtering is introduced as the information expansion path to make the output features more structured and improve the accuracy of vascular segmentation. The proposed method is used to evaluate the performance of the abdominal artery segmentation. The experimentai results show that, compared with the basic network U-Net, the sensitivity and intersection-over-union QIOU) of the proposed method are increased by 2. 84% and 1. 19%, respectively. Compared with the network CE-Net, the sensitivity and IOU are improved by 1. 34% and 1. 61 %, respeclively. [ABSTRACT FROM AUTHOR]

Learning the semantic information and location information of feature map is essential to produce ideal results in retinal image segmentation. Recently, convolutional neural networks have shown strong capabilities in extracting valid information from feature maps. However, convolution and pooling operations filter out some useful information. This paper proposes a new skip attention guided network (SAG-Net) to save feature map's semantic and location information and guide the expansion work. In SAG-Net, the Skip Attention Gate (SAtt) module is first introduced, which is used as a sensitive extension path to pass the semantic information and location information of previous feature maps, not only helps eliminate noise, but further reduces the negative effects of the background. Secondly, the SAG-Net model is further optimized by merging image pyramids to preserve contextual features. On the Drishti-GSl dataset, the joint disc and cup segmentation task proves the effectiveness of our proposed method. Comprehensive results show that the proposed method is superior to the original U-Net method and other recent methods for disc and cup segmentation. [ABSTRACT FROM AUTHOR]

Recognizing age and gender based on facial images is one of the current hot spots in artificial intelligence research. This paper proposes a hybrid model that integrates Convolution Neural Network (CNN), Squeeze-oxidation Network (SENet) and Extreme Learning Machine (ELM). The convolutional layer in the model is used to extract facial features from the face image, the SEnet layer is used to optimize the features extracted by the convolutional layer, and the error minimization extreme learning machine (EM-ELM) is used as a classifier to realize the age and gender recognition of facial ima¬ges. Compared with the existing popular modes, the proposed model adopts the CNN+SENet architecture so that it can extract more representative and optimal feature maps from facial images, and the ex¬tremely fast calculation of EM-ELM makes the model faster and more efficient. Experimental results on multiple unrestricted face datasets show that the proposed model has higher recognition accuracy and speed than other recent related models based on deep earning. [ABSTRACT FROM AUTHOR]

In view of the shortcomings such as low efficiency and easy miss detection when manually selecting moldy tobacco leaves online, this paper proposes a method for screening, classifying and recognizing moldy tobacco leaf images based on convolutional neural network model. Firstly, the tobacco leaf data set is built up. Secondly, a convolutional neural network model is built to initially extract features, screen and extract the main features, then summarize the features of each part, and finally realize the image classification, thereby achieving fast and accurate identification of moldy tobacco leaf images and normal tobacco leaf images. The experimental results show that compared with the artificial moldy tobacco leaves selection method and the traditional image classification algorithm of tobacco leaves, the convolution neural network not only has a high recognition accuracy, but also simplifies the complex process of artificial extraction of image features. [ABSTRACT FROM AUTHOR]

Aiming at the linear bottleneck structure of MobileNetV2, we study the design scheme of the dedicated processor chip. Based on the Layer Fusion mode and the configurable block structure, we design a pipeline structure for bottleneck convolution as well as a corresponding analysis framework. Then, a design space is proposed accordingly, and a software simulator is used to traverse and compare the performance of various schemes in this space. We then find the rules for optimal parameter selection. The validity of the conclusion is verified by hardware behavior simulation. The study can help system chip designers to select or design a suitable MobileNetV2 processor IP design scheme based on their own resource constraints and performance requirements. The paper also provides some inspiration for future automatic processor design. [ABSTRACT FROM AUTHOR]

At present, the neural network model plays an important role in entity relationship extraction tasks. Features can be automatically extracted by a convolutional neural network, Tut it is limited because a fixed window size convolution kernel in a convolutional neural network is used to extract contextual semantic information of words in a sentence. Therefore, this paper proposes a new relational extraction method fusing self-attention and convolutional neural network. The original word vector is calculated by the self-attention mechanism to obtain the relationship between the words in the sequence. The input word vector expresses richer semantic information, which can make up for the deficiency of the automatic extraction features of the convolutional neural network. The experimental results on the SemEval-2010 Task 8 dataset show that, after adding the self-attention mechanism, our model is beneficial to improve the entity relationship extraction effect. [ABSTRACT FROM AUTHOR]

Recurrent neural networks play an important role in sequence recommendation. However, in recommendation, the users behavior sequences are far more complicated than the sentences in natural language processing or images in computer vision. A single recurrent neural network structure is difficult to fully mine user preferences, so this paper proposes a new sequence recommendation algorithm that takes into account both the time information and content information of the sequence. This algorithm is mainly divided into two parts： improved item embedding and sequence preference learning. Firstly，an item embedding method that incorporates a knowledge graph is used to generate high-quality item vectors. Secondly，a sequence modeling method combining convolutional neural networks with long-term and short-term memory neural networks is proposed. Furthermore, an attention-based framework is proposed, which dynamically combines users points of interest. This algorithm is compared with the traditional methods and the existing advanced methods of the same type on the public data set MovieLenslOM. The experimental results show that the average reciprocal ranking MRR@N of the recommended evaluation index and the recall rate Recall@N is improved significantly, which verifies the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Apple is one of the main producing fruits and the main economic crops in many areas. Detecting and grading apples through the image of apple trees under natural environment is helpful to promote the modernization of fruit industry. Combining deep learning with traditional methods, this paper proposes a fruit detection and grading method combining color and apple diameter. In order to improve the detection rate of unobvious targets and the precision of bounding boxes when illumination or fruit coloration is uneven, the convolutional neural network is used to construct an apple detection model and detect apple on feature maps of two scales, b货，(1. 8b-- L*) color components of the image in bounding boxes in CIELAB color space are extracted, the image is binarized, and the target contour is accurately extracted to correct the bounding boxes. Experimental results show that the precision is 91. 60% and the Fl-score value is 87. 62%. According to the image and actual size mapping method, the apple diameter is calculated to achieve the apple grading. Experimental results show that the grading accuracy is 90%. [ABSTRACT FROM AUTHOR]

Traditional detection networks have always had problems of low detection efficiency and low accuracy in complex backgrounds. Aiming at the above problems, this paper further designs the MT-Siam network based on the MTCNN algorithm, which mainly provides accurate posit ion positioning for the independent segmentation of single target, image processing of single target and other operations in the future, so as to quickly obtain the target position and achieve the purpose of improving the detection efficiency. In the experiments, a comprehensive comparison is made on the basic model of YOLOv3, SSD300 and MTCNN to verify the superiority of MTCNN network. The comparative experiments show that, compared with MTCNN, the proposed MT-Siam algorithm can improve the detection speed by 70% to 85% while maintaining high precision. [ABSTRACT FROM AUTHOR]

Due to the light absorption and scattering effect underwater? the quality of underwater images is reduced, which greatly limits the visual range of underwater images. The robustness and accuracy of complex underwater scenes make feature extraction and matching a challenging task. In order to get better registration of underwater images, this paper proposes an underwater image feature registration method based on improved CNN-RANSAC. Firstly, the underwater image is enhanced by the image enhancement algorithm based on deep convolutional neural network? and the VGGNet-16 framework is trained by transfer learning through the underwater image classification datasets, and the improved framework is applied to carry out feature extraction and generate robust multi-scale characteristic descriptor and feature points. After coarse feature matching and dynamic interior point selection, the improved RANSAC algorithm is used to eliminate false matching points. The feature extraction and feature matching experiments are carried out on a large number of underwater image datasets. Compared with the traditional SIFT and SURF registration algorithms? the proposed algorithm can detect more feature points and achieve a significant improvement in matching accuracy. [ABSTRACT FROM AUTHOR]

In view of the shortcomings of YOLO object detection algorithm in small object detection, and the difficulty of achieving real-time performance on embedded platforms, this paper designs an improved YOLO object detection algorithm, called dense_YOLO. The algorithm contains two phases: feature extraction phase and object detection regression phase. In the feature extraction phase, based on the idea of DenseNet structure, a new slim-densenet feature extraction module based on deep separable convolution is designed, which enhances the transmission of small object features and reduces the parameter quantity to accelerate the network propagation speed. In the object detection stage, the idea of adaptive multi-scale fusion detection is proposed to fuse the extracted features, and the objects are classified and regressed on different feature scales, which improves the detection accuracy of small objects. Experimental results show that, compared with the original YOLO object detection algorithm, the dense_YOLO object detection algorithm improves rnAP by 7%, decreases the single picture detection time by 15 ms, and reduces the model size by 90 MB. [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]