Search

Objectives: The quality of network structure is the key to obtain excellent classification performance. Most of existing deep models are designed by the trial-and-error method. When the same model processes different hyperspectral images (HSIs), different network hyperparameters need to be set artificially to adapt to the data characteristics of different images, greatly increasing the difficulty of network design. Aiming at the problem of deep learning model design in hyperspectral image classification, this paper proposes an automatic network structure search method for HSI classification. Methods: In the proposed method, the differentiable structure search technology is first used to search the network structure on the source HSI data sets, then the deep network model is constructed by stacking cells, and finally the classification performance is evaluated on the target HSIs. It should be pointed out that, the proposed method only conducts one search on the source HSI data sets, and the obtained deep model can be applied to other target HSI classification tasks, which can effectively improve the utilization rate of the model. To improve the generalization ability and classification accuracy of the model, multi-source and multi-resolution HSIs are employed to construct the precollected data set, and the partial channel connection operation is introduced to further improve the search efficiency. Results: Three HSIs inculding Houston, Chikusei and Pavia Center are selected as source data. Four HSIs including University of Pavia, Indian Pines, Salinas and Houston 2018 are selected as the target HSIs. The source data is used to determine the network model and the target HSIs are used to evaluate the classification performance. Experimental results show that the proposed method can automatically search out a universal deep network model suitable for HSI classification task, and this model can achieve better classification performance than the conventional deep learning models. The overall classification accuracies of University of Pavia, Indian Pines, Salinas and Houston 2018 are 98.15%, 98.74%, 97.30% and 74.47%, respectively. Conclusions: Compared with other deep models, the model in this paper is automatically determined by differentiable architecture search algorithm according to the data characteristics of HSIs, effectively improving the classification accuracy of target HSIs while avoiding complex network design. Extensive experiments verify the applicability and effectiveness of the proposed method in HSI classification. [ABSTRACT FROM AUTHOR]

Since the objects in remote sensing images have the characteristics of arbitrary direction, dense distribution and large scale differences, object detection in remote sensing images has become a challenging problem. Aiming at this problem, this paper systematically reviewed the work related to rotating object detection in deep learning remote sensing images in recent three years. Firstly, this paper introduced the representation method and characteristics of rotated bounding box. Then, it analyzed the existing methods of rotating object detection in remote sensing images from four aspects: the feature extraction network, the generation of rotating anchor and candidate bounding box, the label allocation and sampling strategy, and the loss function. Next, it introduced the commonly used remote sensing image datasets for rotated object detection, and compared the performance of different algorithms. Finally, this paper prospected rotating object detection in remote sensing images. [ABSTRACT FROM AUTHOR]

To improve the problem that existing stereo matching algorithms have difficulty in maintaining a good balance between higher parallax estimation accuracy and faster model inference speed, this paper proposed an efficient and accurate multi-layer cascaded recurrent stereo-matching network. Firstly, this paper designed a multilayer network to introduce position encoding and self-attention mechanisms on the higher resolution feature maps, then updated the disparity values using the disparity refinement strategies of hierarchical recurrent refinement, cascaded refinement, and recurrent refinement to recover the detailed information of stereo images better. In addition, this paper improved the method of disparity iterative update by using a light- weight group correlation layer at a low scale and an adaptive group correlation layer at a large scale to update the disparity values, which reduced the computational effort of parallax iterative update and improved the inference speed of the model. The experimental results show that the proposed algorithm has a faster model inference speed than others while achieving a competitive parallax estimation accuracy. [ABSTRACT FROM AUTHOR]

Underwater image quality degradation has seriously hindered the development of marine engineering technology. Aiming at the problems of underwater image color distortion, low contrast and uneven brightness, in order to solve the special underwater imaging environment cause the above degradation problems, and facilitate the subsequent research work based on underwater images, this paper proposed an underwater image enhancement network model based on self-calibrated illumination framework. The network model integrated self-calibrated illumination module on the basis of convolutional neural network, which could brighten images quickly and flexibly. The method established a cascaded lighting learning process with weight sharing in the self-calibrated illumination module to process underwater images. Considering the computational burden of cascaded pattern, the method also constructed a self-calibrated module to realize the convergence of the results of each stage. This self-calibrated module reduced the computational cost greatly and realized the automatic adjustment of underwater image brightness. Finally, this paper analyzed the experimental results qualitatively and quantitatively, and compared eight advanced enhancement algorithms. The results show that the proposed method is obviously superior to other compared algorithms. The results of the proposed method improve the contrast, correct the color, improve the uneven brightness, and enhance the visual effect. The qualitative and quantitative analysis results show the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Most of the Encryption-then-Compression (ETC) methods can only obtain several limited fixed compression ratios. However, arbitrary encryption compression ratios instead of limited fixed compression ratios are more suitable for practical requirements. To this end, this paper proposed a lossy compression algorithm for encrypted color images with arbitrary compression ratios. It combines the uniform and random downsampling to compress the encrypted images, obtaining arbitrary compression ratios of an encrypted image. The receiver receives the compressed sequence of the encrypted image and obtains the decrypted image by decompression and decryption. The proposed scheme then characterizes the lossy reconstruction of the original image from the decrypted image as an optimization problem with the downsampling compression-based constraint. This scheme designs a convolutional neural network-based image reconstruction model for lossy ETC to resolve this problem, which is denoted the ETC-oriented Reconstruction Network (ETRN). ETRN consists of shallow feature extraction (SFE), residual in residual (RIR), residual content supplementation (RCS), and down-sampling constraint (DC). The experimental simulation results show that the encrypted color image lossy compression algorithm proposed in this paper can obtain excellent compression and reconstruction performance, which fully demonstrates the feasibility and effectiveness of this method. [ABSTRACT FROM AUTHOR]

Although high-quality and high-resolution depth maps can significantly improve the performance of computer vision tasks in various natural scenes, the limitations of depth camera hardware make the depth maps captured by consumer-level depth cameras have problems such as low resolution, poor quality, invalid holes, etc. Depth super-resolution reconstruction (DSR) is a kind of technology that can effectively improve the resolution and quality of depth maps, and DSR has become a research hot-spot in the field of computer vision. Firstly, this paper introduced the definition of DSR and the research progress of DSR algorithm at home and abroad in recent years, and then mainly stated and analyzed deep learning DSR reconstruction algorithms. Next, it introduced the depth image quality evaluation criteria. Finally, this paper prospected the application fields of DSR and the challenges as well as the opportunities in the future. [ABSTRACT FROM AUTHOR]

Aiming at the problem that the existing single image super segmentation reconstruction methods are difficult to capture complete and effective information in the image, this paper proposed a joint image-frequency supervised image super-resolution reconstruction algorithm, which aimed to utilize the complementary information between multiple domains, thereby obtaining a more complete image feature representation. Firstly, by analyzing the characteristics of image pixels after converting to frequency domain space, this paper proposed a new frequency domain distance supervision loss according to its complex representation, which effectively applied spectral information to the optimization process of convolutional neural networks. Based on the characterization characteristics of different frequency bands in the domain, this paper constructed the spectral weighting loss on the basis of the distance loss in the frequency domain, and applied it to the two frequency bands of low frequency and high frequency respectively. Finally, combined with the supervision of the image domain, it formed a joint optimization of multiple domains and got good performance. The results on Setl4, BlOO and Kodak show that the PSNR and SSIM of the proposed algorithm reach 33.47 dB and 0.985 9, respectively, and the proposed algorithm achieves the best performance compared with several image super-resolution methods. [ABSTRACT FROM AUTHOR]

[Objectives] Under the high-density breeding mode, broilers in cages are prone to respiratory diseases, which are difficult to prevent. In this paper, a cough detection method for whole-life cycle cough of white-feathered broiler based on vocalization characteristics and deep learning was designed to automatically monitor the cough vocalization of broilers and provide early warning information in time. [Methods] Digital filtering, spectral subtraction denoising and endpoint detection were performed on the vocalization signals of broilers at three different growth stages: 10-19 days old, 20-29 days old and after 30 days of age, and the filter bank (FBank), Mel-frequency cepstral coefficients (MFCC) features and their respective first-order and second-order differences were used as the input of the VGG16 and ResNet18 neural networks to complete the three-category model training of cough, tweet and other vocalizations. [Results] The recognition models constructed by using different vocalization features and neural networks in each age group could accurately realize the vocalization classification, and the FBank-VGG16 model was better at 10-19 days old and 20-29 days old, and the accuracy rates were respectively 94.29% and 97.65%. The accuracy rate of the MFCC-ResNet18 model after 30 days was higher than other models at 98.66%. As the age increased, the overall recognition accuracy of each model increased, ranging from 3% to 7%. [Conclusions] This experimental method could quickly and accurately identify the cough sounds of broilers at different life stages in the actual production environment, and provided technical support for the early detection of respiratory diseases in caged broilers. [ABSTRACT FROM AUTHOR]

To solve the problem of combustion state recognition in solid oxide fuel cell (SOFC) combustor, this paper proposed a combustion state recognition method based on attention mechanism and image feature pyramid. The method adopted the adaptive gamma correction with weighting distribution (AGCWD) to standardize the input images. It combined two single full connections with 1×1 convolution to replace the squeeze and excitation structure, and proposed a hybrid attention structure combined with spatial attention structure to enhance the ability of feature extraction. To provide multi-scale information communication capability, it constructed the multi-scale bidirectional fusion pyramid by means of bidirectional computation and multi-scale fusion. The experimental results show that the proposed method reaches 99.22% accuracy under the premise of 3.98 M parameters and 397 M floating point operations (FLOPs), and effectively identifies the combustion state in SOFC combustor. [ABSTRACT FROM AUTHOR]

In order to solve the problems of high computational complexity, large computational cost and large number of parameters, this paper proposed the parallel group convolution algorithm based on the domestic SW26010P multi-core processor. The core idea was to use the unique data layout, through the multi-core mapping processing, parallel computing. Experimental results show that compared with single-core serial algorithm, the proposed parallel group convolution algorithm can achieve the highest speed-up ratio of 79.5 and the maximum effective computing power of 186.7MFLOPS. After data parallel optimization of the parallel group convolution algorithm by SIMD instruction, the algorithm obtains the highest speed-up ratio of 10.2 compared with the parallel group convolution algorithm before optimization. [ABSTRACT FROM AUTHOR]

Expert recommendation is one of the research hotspots in the online Q&A community. However, the majority of the currently used algorithms concentrate on matching user's static interest and question information, ignoring the efficient capture of user' s dynamic interest representation information, which results in a lack of recommendation accuracy. In order to address the aforementioned issues, this paper proposed an expert recommendation algorithm for the Q&A community that incorporated multi-head self-attention. Firstly, this algorithm built out question encoders of convolutional neural networks and attention mechanisms to process target question and user' s historically answered questions, in order to extract the corresponding question representations. Secondly, it treated the sequence of user's historically answered questions as a time series and used the multi-head self-attention mechanism to learn the dynamic interest representations hidden in the sequence, and combined with the user' s static interest representations to obtain the user' s comprehensive interest representations. Finally, it calculated the similarity between the target question representation and the comprehensive user representation to generate recommendation results. Based on the real data from the Zhihu Q&A community, the experimental results show that the performance of this algorithm is significantly better than the current popular expert recommendation algorithm. [ABSTRACT FROM AUTHOR]

Mild cognitive decline is the early stage of Alzheimer's disease, and the feature extraction and classification of mild cognitive decline using EEG signals is an important method for diagnosing mild cognitive decline. In the automatic detection technology of mild cognitive decline based on electroencephalogram artificial intelligence, the existing research only extracted a certain feature in the electroencephalogram signal or simply concatenates multiple features, which caused these methods to fail to well consider the correlation between different features and it would cause the problem of dimensional disaster. This paper proposed a convolutional neural network based automatic detection algorithm for resting state electroencephalogram data of mild cognitive decline. By extracting the power spectrum and brain network features of the electroencephalogram, it fused the two features by matrix operation, and designed a convolutional neural network to classify the fused features. This method achieves a high accuracy rate on the data set collected by a hospital in Shanghai. In addition, by inputting different subsets of the feature set, this method found the few groups of features that contribute the most to mild cognitive decline, thereby it also had a certain interpretability. Experiments on the dataset of this paper, it p roves the advantages of the power brain network for the automatic diagnosis of mild cognitive decline. [ABSTRACT FROM AUTHOR]

This paper proposed a natural language sentence matching method that combined high-level and low-level semantic information to solve the problems about current natural language sentence matching method fail to integrate common semantic information and it is difficult to capture deep-semantic information. First of all, the method used pre-trained word vector GloVe and character-level word vector to obtained the word embedding representation of sentence P and sentence Q. Secondly, this paper encode red P and Q with bidirectional LSTM, then it contained low-level semantic information through preliminary fusion of P and Q. Thirdly, this paper calculated bidirectional attention between P and Q, then spliced them together to get semantic representation, afterwards it calculated its self-attention to obtained high-level semantic information. Finally, this paper used a heuristic fusion function to fuse the low-level semantic information with the high-level semantic information to obtain the final semantic representation, and it used a convolutional neural network to prediction answers. This paper evaluated the proposed model on two tasks, such as recognition textual entailment, paraphrase recognition. This paper conducted experiments on the SNLI dataset and the Quora dataset. The results show that the accuracy of the proposed algorithm on the SNLI test set is 87. 1%, and the accuracy of the Quora test set is 86. 8%, which verifies the effectiveness of the algorithm in the task of natural language sentence matching. [ABSTRACT FROM AUTHOR]

Most of the existing steganalysis methods have weak generalization ability and cannot effectively detect unknown steganalysis algorithms, which makes the accuracy of their classification greatly reduced in the practical application process. To solve this problem, this paper proposed an image steganalysis method based on group convolution and snapshot ensembling（Snapshot Ensembling steganalysis network, SENet）. Firstly, residual convolution block and group convolution block are used to extract the features of the image. Secondly, the model with the best performance is obtained as the snapshot model in each training period. Finally, the selected snapshot models are integrated to classify the images. This method uses the techniques of group convolution and snapshot ensembling to avoid the high training cost of traditional integration methods and the limited generalization ability of a single classifier. Experimental results show that this method can improve the accuracy of steganographic analysis model, and can effectively classify when the training set and test set are mismatched, and has high model generalization ability. [ABSTRACT FROM AUTHOR]

This paper proposed a self-encoding salient region extraction algorithm based on multi-layer convolutional to solve the problem that the existing convolutional neural network wasn’t sufficient in extracting image features, which led to the poor effect of salient extraction. When convolutional networks were used to extract image features, the shallow convolution features generally extracted the detailed features of the image, such as color, texture and location features, and the deep convolution features were generally referred to the semantic features of the image. On the coding layer, this algorithm downsampled the shallow convolution features and fused them into the deep convolution features, and upsampled the deep convolution features to the shallow convolution feature. Experiments show that this method can greatly improve the coding quality. In the decoding, it also fused the convolutional features by encoding to obtain the lost information in the decoding and also obtained a better decoded image. In addition, this paper designed a layer-by-layer supervision method to guide the training of the decoding layer, which used the standard region extraction map by down-sampling as the standard image for supervision training. The experimental results show that this algorithm can increase the F-measure by 0.071 on the basis of PAGRN, and reduce the MEA by 0.031 on average. [ABSTRACT FROM AUTHOR]

The currently widely used lightweight neural networks still have the problems such as large number of calculations and parameters, which make cheap mobile devices with low computation resources cannot quickly complete the image classification tasks. Aiming at this problem, this paper proposed a lightweight neural network that was more suitable for using on cheap mobile devices with low computation resources. This paper introduced cheap linear operations and concatenations of feature maps to reduce the number of calculations and parameters of the neural network. Also, this paper introduced improved residual structure, attention mechanism and label smoothing to improve the accuracy of result. Experiments based on PD-38 dataset show that the proposed neural network can achieve higher classification accuracy by using a smaller number of calculations and parameters than traditional lightweight neural networks. Experiments on the public dataset CIFAR-10 further show that the neural network is universal. [ABSTRACT FROM AUTHOR]

The research in 3D human pose estimation from 2D images has achieved great success in recent years. However, the performance of existing 3D human pose estimation methods may degrade significantly in complicated scenarios. To improve the accuracy and robustness of 3D human pose estimation in unconstrained scenarios, this paper proposed ARGP-Pose, a monocular 3D pose estimation framework by exploring the relationship between the joint points of a 3D human pose. To be more specific, the proposed method included a new joint point preprocessing method and a 3D pose estimation network. The preprocessing method enhanced structural features and extracted the relationship among joint points, which were used as input of the following pose estimation network. Also, the proposed network fused local information of each joint point and the global information of the overall pose for rich feature extraction. Additionally, the proposed method extracted the temporal information by using a self-attention module, which achieved further performance boost. Last, for a complex human pose, this method decomposed the prediction of the whole pose into the prediction of each point, which again improved the estimation accuracy for human bodies with complex pose variations. The experimental results obtained on several well-known benchmarking datasets, such as Human3.6M and HumanEva-I, demonstrate the merits and superiority of the proposed method. [ABSTRACT FROM AUTHOR]

Manual, brush or ion blowing methods are adopted in the clearing system of detachable electric energy meter, resulting in low accuracy of stain identification and poor cleaning effect. Therefore, the methods of binocular vision and machine learning were used to improve it. The system takes the control center as the core, the identification module was connected with the cleaning service subsystem through the communication protocol, and the system framework was constructed. A binocular vision system was established. After the binocular camera was calibrated by neural network, the image information of disassembled electric energy was collected, and the stain location was completed through coordinate conversion and parallax calculation. The positioning results were input into the convolution neural network (CNN), and the dirt identification of the disassembled watt hour meter was completed through grid division, feature mapping and other operations. The cleaning device was designed in the cleaning service subsystem, and the electric energy meter was removed according to the stain identification results to complete the design of the cleaning system. Taking cleaning 500 electric energy meters as an example, the experimental results show that the method in this paper takes 22 minutes, the recognition error is about 10%, and the verification effect of the cleaned electric energy meter is improved from 70% to 97%. [ABSTRACT FROM AUTHOR]

In general adversarial attacks oriented to sample sets, the general perturbation design that causes most sample to output errors is the key to the research. This paper takes the typical convolutional neural networks as the research object, summarizes the existing general perturbation generation algorithms, and proposes a general perturbation generation algorithm that combines batch random gradient ascent and spherical projection search. In each iteration of the algorithm, a small batch of samples are extracted from the sample set, and the general perturbation is calculated by using the random gradient rising strategy which reduces the value of the loss function. The general perturbation is then projected to the high-dimensional spherical surface with a radius of ε, so as to reduce the search space of general disturbances. The algorithm also introduces a regularization technique to improve the generation quality of general disturbances. Experimental results show that compared with the baseline algorithm, the attack success rate is significantly increased, and the solution efficiency of general perturbation is improved by about 30 times. [ABSTRACT FROM AUTHOR]

Because convolutional neural networks( CNN) mostly focus on global feature learning and ignore local feature information containing more details, it is difficult to improve the accuracy of indoor scene recognition. To solve this problem, this paper proposed an indoor scene recognition method based on improved global-local attention network ( GLANet). Firstly, it used GIAN et to capture the global and local features of the scene image to increase the detail information of the image features. Then, it introduced a non-local attention module into the local network to further preserve the detail features of the image through the convolution of attention diagram and feature graph. Finally, it used various features of different stages of the network for classification. Through training and verification on the MIT lndoor67 and the SUN397 dataset, the recognition accuracy of the proposed method was increased by l. 98% and 3. 07% respectively compared with the LGN method. Experimental results show that the algorithm can effectively capture global semantic information and fine spatial details, and significantly improves the recognition accuracy. [ABSTRACT FROM AUTHOR]

A single semantic category label cannot describe comprehensively the image content because remote sensing images contain various object categories. The task of multi-label image classification is more challenging. By exploring the depth graph convolutional network(GCN), this paper made up for the lack of relevance of label semantic information in multi-label remote sensing image classification, proposed a new multi-label remote sensing image classification network, multi-label remote sensing image classification network based on the GCN. It contained three parts: image feature learning module, classifiers learning module based on GCN, and image feature differentiating module. Compared with the related models on the public multi label remote sensing datasets planet and UCM, the method can get better classification results on the multi label remote sensing image classification task. The method used modules such as graph convolution to apply multi-label image classification methods to remote sensing, which improved the model classification ability and shortened the model training time. [ABSTRACT FROM AUTHOR]

The existing machine reading comprehension models mainly use the recurrent neural network to process the text sequence information, which easily leads to slow down the training and prediction speed, and low accuracy of model prediction, this paper proposed a span-extraction machine reading comprehension algorithm QA-Reader. The algorithm first used a large pre-trained language model RoBERTa-www-ext to obtain the word embedding representations of questions and text. Secondly, it used deep separable convolution neural network and multi-head self-attention mechanism to encode. Then it calculated the bidirectional attention and the text’s self-attention to fuse the correlation information between text and problem, and splicing to get the final semantic representation. Finally, the model encoder predicted the answer. For the unanswerable questions, the model calculated the unanswerable probability. Experiments were carried out on the Chinese span-extraction machine reading comprehension datasets. The results show that compared with the baseline model, the QA-Reader model has improved EM andF1 values by 3. 821% and 2. 740% in performance, and the training speed has improved by 0. 089%. [ABSTRACT FROM AUTHOR]

The size of the receptive field of the convolutional neural network is related to the size of the convolution kernel. And the traditional convolution uses a fixed-size convolution kernel, which limits the feature perception ability of the network model. In addition, due to the parameter sharing mechanism of the convolutional neural network, it used the same feature extraction method for all pixels in the spatial region. However, there are differences in the distribution of noise signals and clean speech signals in the noisy spectrogram, especially in the complex noise environment, the general convolution method is difficult to achieve high-quality speech signal feature extraction and choosing. In order to solve the above problems, this paper proposed a multi-scale region adaptive convolution module, which used multi-scale information to improve the feature perception ability of the model and automatically allocated the area adaptive convolution achieve and improved the denoising ability of the model. The experiments on the TIMIT public datasets show that the proposed algorithm has achieved satisfactory results in the metrics of speech quality and intelligibility. [ABSTRACT FROM AUTHOR]

Due to the small amount of microscopic features image data of traditional Chinese medicinal materials powder, unbalanced distribution of sample classes and small difference between classes, it is difficult to achieve a satisfying classification effect through traditional image classification methods. To solve the above problems,this paper proposed an improved method of deep convolution neural network based on dynamic ReLU and attention mechanism model. Firstly, it used Xception as the basic network, which had an obvious effect on small sample data classification. Secondly, it replaced the static ReLU activation function in the network with the improved dynamic ReLU function, so that each sample had its own unique ReLU parameters. Finally, it embedded the improved SE module in the network to enable the network to learn the importance of each feature channel automatically. The proposed method can make the network pay more attention to the detailed information in the image, and can solve the problem of unbalanced distribution of sample classes and small differences between classes. The experimental results show that the image classification accuracy of 56 kinds of traditional Chinese medicinal materials powder vessel is increased by about 1.5% to 93.8%, which demonstrates that the proposed method is advantageous over other image classification methods. [ABSTRACT FROM AUTHOR]

User or item side information is widely used in numerous studies to alleviate the problems of data sparseness and cold start in video recommendation, which proved to be effective to some degree. However, the key part of auxiliary information was neglected in previous studies. This paper conducted research on this problem and proposed a deep hybrid recommendation model that incorporated dual attention mechanism. This model mined the hidden vectors on the item side through the convolutional neural network integrated with the self-attention mechanism and extracted the hidden vectors on the user side through the autoencoder fused with the self-attention mechanism. In this way, vital information on both the item side and user side was deeply mined, which was then combined with the classic probability matrix decomposition to realize score prediction. A large number of experimental results in two public datasets show that the results of the proposed model improves a certain extent compared with baseline models such as ConvMF+,PHD, and DUPIA. [ABSTRACT FROM AUTHOR]

In order to recognize the knots accurately and reduce the waste of wood, a method of recognizing the knots based on GoogLeNet convolution neural network based on the deep learning framework of Pytorch was proposed in this paper. This method used GoogLeNet network to extract features automatically from RGB images of seven kinds of spruce knots, such as decayed knots, dry knots and encased knots et al, so that classification and recognition can be realized without image preprocessing. Global average pooling was used to replace the fully connected layer, and the network parameters can be reduced. At the same time, in order to prevent overfitting, Dropout mechanism was used in the network. The experimental results showed that, the recognition accuracy of seven kinds of knot defects using the convolutional neural network can reach 95. 42%, and the GoogLeNet can identify the knots accurately and effectively in the image processing of wood knot defects. [ABSTRACT FROM AUTHOR]

Facial features play different functions in expressing different emotions, the only use of convolutional neural network for feature extraction and expression recognition of facial features may not lead to enough focus on the key feature information of facial expressions. However, the feature information only extracting from the key parts such as eyes or mouth may not be sufficient. This p aper proposed a facial expression recognition method of tri-path networks based on the optimal convolutional neural network called TP-FER. Based on the convolutional neural network, this method used three input channels to focus on the face, eyes and mouth regions respectively for feature extraction and expression discrimination. Then it used the fusion technology based on the decision layer to make the final decision. At last, it adopted relative majority voting method based on the recognition results of the three channels to obtain the overall optimal recognition rate. This paper conducted several experiments on JCK + dataset and self-built dataset with this method. The results show that the ove rall expression recognition rate is improved in both datasets. This method can make full use of the different channel characteristics that focus on not only the key features such as eyes and mouth but also the comprehensive information of the whole face, and improve the express ion recognition rate effectively. In addition, it can also provide data support for neuropsychological related research. [ABSTRACT FROM AUTHOR]

Remapping the pixel intensity range of the high data depth computed tomography( CT ) image to an observable range for window technology will lose image information, and the method of automatic segmentation after only one pre-processing of CT images is not well adapted to the tumor segmentation. This paper presented a two-step optimized intensity windowing preprocessing method for segmented CT values. It used the histogram of CT value and window technology of liver and its tumors to preprocess the original and liver image respectively, improved the contrast of the area of interest. Compared the automatic segmentation accuracy of liver and tumor with different intensity windows in the cascaded convolutional neural network (CNN), obtained the optimal CT range ofliver and tumor, which were [ -40, 190] HU and [ - 25, 145] HU, respectively. This method was tested on the 3 DIRCADb dataset with three mainstream segmentation networks, UN et, ResN et, and FCN. It proves that the influence of different CT value ranges on liver segmentation accuracy can reach 6. 75 %, and the segmentation accuracy of different networks can be improved by about 2%. [ABSTRACT FROM AUTHOR]