ABSTRACT. Due to the scattering and absorption of light, underwater images typically suffer from severe color distortion and contrast degradation. To address the above challenges, we propose a color correction and adaptive contrast enhancement for underwater image enhancement. In our work, we first design the dedicated fractions to compensate the lower color channels, which is calculated by considering the ratio of the differences between the upper and lower color channels to the lower color channels. Subsequently, an adaptive contrast enhancement method is applied to each color channel is proposed to produce the background-stretched and foreground-stretched images, the combination of these two images significantly improves the contrast of the output image. Finally, a unsharp masking technique is applied to sharpen the overall image. Experiments on the standard underwater image enhancement benchmark dataset have proven that proposed algorithm can output high-quality underwater images with high contrast, good visibility and natural appearance, even outperform than other state-of-the-art underwater image enhancing algorithms. In addition, our algorithm can also be well extended to other types of degraded images, such as hazy and weakly illuminated images.

ABSTRACT. Recent years, with the improvement of human life, people's requirements for food quality are increasing. People need an eective food recognition algorithm to get more useful food informations. For the problems of food recognition algorithm which are low accuracy and slow recognize speed. This paper proposes a food recognition algorithm based on CNNs (Convolution neural networks). First of all, the proposed algo- rithm preprocess the food images which are collected from internet. And then use the traditional convolution extract the features from food im- ages. The jumping convolution which is designed in this paper to extract food images feature jumping and combines the feature from traditional convolutions. This food recognition algorithm can solve food recognition issues eectly, in addition to reduces the computation parameters. The experimental results compared with other deep learning networks, and the proposed algorithm gets a good result, which can recognize the food quickly and reduce the training time.

ABSTRACT. Due to the nonlinear and underactuated characteristics of unmanned surface vehicle (USV) system and the uncertainty of environmental model, it is hard to establish accurate dynamic model and control law obtained by traditional algorithm which is too complex and has no engineering practice realization. In this paper, based on deep reinforcement learning algorithm of deep deterministic policy gradients (DDPG), the line of sight (LOS) algorithm is used firstly to obtains the expected value of heading angle of USV according to the current time position and the expected trajectory of USV. Meanwhile, we adopt the double Gaussian reward function to evaluate the training action, so as to obtain the optimal control action to realize the accurate tracking control. Finally, compared with explicit model predictive controller (EMPC) and linear quadratic regulator (LQR), the designed track controller based on DDPG has shorter adjusting time and smaller overshoot than EMPC and LQR.

ABSTRACT. The current description language used to describe auction-based market is only capable of representing and reasoning the basic rules of the auction in one certain auction scene, while the strategies of players are not fully taken into consideration. In order to improve the description ability of language, the Epistemic Description Language for Auction is proposed by using epistemic strategy to represent and reason more domain knowledge in auctions. Firstly, the language is extended by epistemic operators and semantics. Secondly, the Auction Description Language with Extension is used to represent and the economic properties and all the properties are formalized. Finally, the extended language is proved through conducted experiments that strategies for better playing are considered, and more domain knowledge of different auction-based markets can be represented.

ABSTRACT. In this paper, we construct a deep learning model based Multilayer Perceptron (MLP) to discover the brine formation mechanism. Brine data are divided into three data sets with the same six chemical substance features. The training and validation sets are collected from the published data, while the test set was extracted from various geographic positions and depths in the Laizhou Bay, China. Based on the validation set, the performance of this deep learning model is compared with three state-of-the-art prediction algorithms using six evaluation metrics. The experimental results show that the proposed model outperforms these three algorithms. Specifically, the proposed method has around 8% in precision improved to the decision tree algorithm. Furthermore, the trained model is utilized to predict the real-world test set; the results of which confirm the existence of frozen brine for the first time.

ABSTRACT. Video summarization is one of the challenging tasks in the field of multimedia content analysis. In prior works, video summarization is usually formulated as a sequential decision-making process, in which the input is a sequence of video frames and the output is a subset of the original frames. While Long Short-Term Memory (LSTM) is good at temporal dependencies modeling, it can only handle short videos of up to 80 frames in length. In this paper, we propose a novel framework named Attention-based Hierarchical LSTM networks for Video Summarization (AHVS) that integrates delicate feature extraction, temporal dependencies modeling and video summary generation. Specifically, 3D CNN is employed to extract spatial-temporal features and an attention-based hierarchical LSTM module is designed to capture the long-range temporal dependencies. In addition to advances in modeling techniques, we treat video summarization as an imbalanced class distribution problem and design a cost-sensitive loss function to train our model. Extensive experiments on two popular benchmark datasets, i.e., SumMe and TVSum, show that our proposed method achieves competitive results comparing with other state-of-the-art approaches.

ABSTRACT. Transmission line is related to the security and stability of the power system, and it is the basis of the ensuring power supply. It is very important to ensure the normal operation of the transmission line and timely find and repair the faulty transmission line. This requires accurate identification of the fault type and the distance when a fault occurs. Therefore, this paper comprehensively summarizes the transmission line fault diagnosis technology and its research status. And the fault analysis method, traveling wave method, intelligent positioning method and other methods are analyzed and combed. On this basis, the research and application prospects of the artificial intelligence positioning method based on deep learning are summarized and prospected.

ABSTRACT. Vehicle detection is one of the most fundamental aspects of traffic surveillance systems. However, the shadow problem often hinders the accuracy of vehicle detection. Shadows are often mistakenly understood as the parts of a vehicle, causing objects loss, or shape distortion. Collected data with these errors are unreliable, therefore shadow detection and elimination is a key step to promote the accuracy of vehicle detection. This paper proposes a robust vehicle detection method with shadow elimination. The method is divided into two steps: Firstly, we extract foreground regions using a background differential method based on edge information, then we detect and eliminate the shadows from the foreground regions combined with grayscale information, edge information and prior knowledge. Compared with the deep learning methods, such as faster R-CNN, and SSD, experimental results demonstrate that the proposed method improves the accuracy of vehicle detection more than 10%.

ABSTRACT. Aiming at the problem of small target detection, this paper first sorts out the development and current situation of target detection algorithms, and systematically summarizes the research progress of target detection algorithms on complex ground backgrounds. Secondly, we start with two major categories of hyperspectral small target detection and infrared small target detection, and each category is analyzed from different methods. Then we take the representative algorithm as an example to analyze its detection performance and its application under the actual complex ground background conditions. Finally, we respectively make prospects and predictions for each type of algorithm in the application of complex ground background target detection, which provides a reference for future research on small target detection problems.

ABSTRACT. Seed identification exhibit potential for ensuring the healthy development of agriculture and food security. In this study, in order to detect corn kernel varieties quickly, accurately and nondestructively, near infrared spectroscopy (NIRS) was employed to develop identification methods for corn seeds. Four varieties of corn kernels were evaluated. Classification models based on support vector machine (SVM), support vector machine-genetic algorithm (SVM-GA), BP neural network, and principal component regression (PCR) were developed. The experimental results show that SVM-GA has the best detection effect, and the accuracy is 97.5%. This study showed that combined NIRS and SVM-GA model could be an effective method for the classification of corn seeds.

ABSTRACT. Facial landmark detection is a key research content in the field of facial recognition analysis. It is also a challenging task for facial image acquisition under natural field conditions, which is subject to factors such as illumination, facial expressions, occlusion, and large angle pose. This paper mainly aims at the high error rate of facial landmark detected by the TCNN network model at one time. A serialization-based facial landmark detection algorithm is proposed. The facial landmarks are detected in the order from salient landmarks to sub-salient landmarks and then from sub-salient landmarks to non-salient landmarks. And then the detected salient landmarks are feature-coded and fused with the original image to predict the position of the non-salient landmarks, which effectively solves the problem of high detection error caused by non-salient facial landmarks without obvious features. The proposed Sequenced Tweaked CNN(STCNN) architecture is shown to outperform compared landmark detection methods in an extensive battery of tests on the AFW and AFLW benchmarks. Finally, in order to promote the reproducibility of our results, we make code and trained models publicly available at https://github.com/haiduo/STCNN.

ABSTRACT. In this paper, we combine the embedded deep pyramid network and the deep FM recommendation model to complete the recommendation of proprietary Chinese medicines. Using real electronic medical records provided by Beijing University of Chinese Medicine and the Institute of Computer Science, Chinese Academy of Sciences as data sources; according to the patient’s main complaint, symptoms, tongue, pulse, age, gender, and other characteristics; using a deep pyramid network based on embedded patient feature fusion to predict the patient’s card Candidate categories, realize long-distance text association through text region embedding and fixed number of feature maps; and use deep FM model to process the sparse and dense features of the patient with the FM layer and the DNN layer separately, and finally real-time features and predictions for the patient The syndrome types are sorted. After comparative experiments and training, the AUC index is 78.90%, and the ACC value is 83.26%. Its overall performance is better than other RNN, CNN, and fast text network models. The experimental results show that the recommendation algorithm of Chinese patent medicine based on DPCNN-DeepFM has certain reference significance for the auxiliary clinical diagnosis of patients.

ABSTRACT. Professionally broadcasted sports videos usually have long durations but contain only a few exciting events. In general, professional bodies and amateur content creators spend thousands of man-hours to manually crop the exciting video segments from these long-duration videos and generate handcrafted highlights. Sports enthusiasts keep them updated with the latest happening based on such highlights. There exists a need for a method that accurately and automatically recognizes the exciting activities in a sports game. To address this issue, we present a deep learning-based network SPNet that recognizes exciting sports activities by exploiting high-level visual feature sequences and automatically generates highlights. The proposed SPNet utilizes the strength of 3D convolution networks and Inception blocks for accurate activity recognition. We divide the sports video excitement into views, actions, and situations. Moreover, we provide 156 new annotations for about twenty-three thousand videos of the SP-2 dataset. Extensive experiments are conducted using two datasets SP-2 and C-sports, and the results demonstrate the superiority of the proposed SPNet.

ABSTRACT. Accurate lung tumor segmentation plays an important role in radiotherapy and targeted therapy. Towards this end, positron emission tomography (PET) and computed tomography (CT) provide mutually-compensative evidences. PET clearly recognizes the lung tumor, while the tumor boundary is blurred in such a modality. By contrast, CT has a better resolution but a lower contrast between the tumor and normal tissues. As the result, accurate lung tumor segmentation is challenging by either one single imaging modality. It is well known that combining PET and CT imaging evidences benefits lung tumor delineation. However, the inconsistence between PET and CT poses difficulties in fusing these two modalities. Aiming to solve this problem, a novel joint level set model is proposed to combine the evidences of PET and CT in a united energy form, achieving a co-segmentation in these two modalities. The convergence of the co-segmentation model corresponds to the most optimal tradeoff between the PET and CT. The different characteristics in these two imaging modalities are considered in the adaptive convergence process which starts mostly with the PET evidence to constrain the tumor location and stops mostly with the CT evidences to delineate boundary details. The adaptability of our proposed model is automatically realized by stepwise moderating the joint weights during the convergence process. The performance of the proposed model is validated on 20 nonsmall cell lung tumor PET-CT images. It achieves an average dice similarity coefficient (DSC) of 0.846±0.064 and positive predictive value (PPV) of 0.889±0.079, demonstrating the high accuracy of the proposed model for PET-CT images lung tumor co-segmentation.

ABSTRACT. The Drosophila visual system is extremely sensitive to moving targets, which provides a wealth of biological inspiration for the research of target motion perception in complex scenes, and also lays a biological theoretical foundation for the establishment of artificial drosophila visual neural networks. Drosophila's vision has been extensively studied in physiology, anatomy, and behavior, but our understanding of its underlying neural computing is still insufficient. In order to gain insight into the neural mechanism in Drosophila vision and take better advantage of its superiority in motion perception, we propose a Drosophila vision-inspired model, which constructs a complete Drosophila visual motion perception system by integrating continuous computing layers. Our hybrid model can fully demonstrate the motion perception process in Drosophila vision. In addition, the Drosophila vision-inspired model can also be exploited to salient object detection in dynamic scenes. This novel salient object detection model is different from the previous in that it can accurately identify the motion of interest (MOI) while suppressing background disturbances and ego-motion. Comprehensive evaluations using standard benchmarks demonstrate the superiority of our model in salient object detection compared with the state-of-the-art methods.

ABSTRACT. Predicting the population density in certain key areas of the city is of great importance. It helps us rationally deploy urban resources, initiate regional emergency plans, reduce the spread risk of infectious diseases such as Covid-19, predict travel needs of individuals, and build intelligent cities. Although current researches focus on using the data of point-of-interest (POI) and clustering belonged to unsupervised learning to predict the population density of certain neighboring cities to define metropolitan areas, there is almost no discussion about using spatial-temporal models to predict the population density in certain key areas of a city without using actual regional images. We abstract 997 key areas in Beijing and their regional connections into a graph structure and propose a model called Word Embedded Spatial-temporal Graph Convolutional Network (WE-STGCN). WE-STGCN is mainly composed of three parts, which are the Spatial Convolution Layer, the Temporal Convolution Layer, and the Feature Component. Based on the data set provided by the Data Fountain platform, we evaluate the model and compare it with some typical models. Experimental results show that the Spatial Convolution Layer can merge features of the nodes and edges to reflect the spatial correlation, the Temporal Convolution Layer can extract the temporal dependence, and the Feature Component can enhance the importance of other attributes that affect the population density of the area. In general, the WE-STGCN is better than baselines and can complete the work of predicting population density in key areas.

ABSTRACT. Recently, boundary information has gained more attention in improving the performance of semantic segmentation. This paper presents a novel symmetrical network, called BASNet, which contains four components: the pre-trained ResNet-101 backbone, semantic segmentation branch (SSB), boundary detection branch (BDB), and aggregation module (AM). More specifically, our BDB only focuses on processing boundary-related information using a series of spatial attention blocks (SABs). On the other hand, a set of global attention blocks (GABs) are used in SSB to further capture more accurate object boundary information and semantic information. Finally, the outputs of SSB and BDB are fed into AM, which merges the features from SSB and BDB to boost performance. The exhaustive experimental results show that our method not only predicts the boundaries of objects more accurately, but also improves the performance of semantic segmentation.

ABSTRACT. In some dangerous environment, it is difficult to complete required tele-operation tasks. In this paper, we propose a multilateral tele-operation control, based on nonlinear disturbance control method. The stability is analyzed and numerical simulations verify the effectiveness of the proposed method.

ABSTRACT. Due to the long calculation time of task allocation algorithms and the long flight distance of the UAV during task allocation, this paper proposes a multi-UAV task allocation method based on three bionic algorithms, namely the ant colony algorithm, the bat algorithm, and the gray wolf algorithm. The TABA (Multi-UAV Task Assignment Method Based on Bionic Algorithms) method dynamically allocates UAVs based on the number of task points and includes a comparison mechanism. The results of experiments demonstrate that, among the three algorithms, the proposed TABA method reduces the UAV distance traveled by about 30% on average, and the UAV travel time is about one-third that of the original algorithm on average, thereby revealing the effectively reduced algorithm complexity. Moreover, the algorithm can reduce the travel time of UAVs.

ABSTRACT. The Software Defined Networks is a popular network architecture that separates control and data plane. It is completed as a software to move control plane functions from network hardware to a controller node. To optimize the consistency of the network, a controller placement problem (CPP) could be treated as an optimization problem to minimize the latency between nodes by setting their location. In this paper, as a modern Evolutionary Computation Algorithm, grey wolf optimization algorithm recently has been successfully applied into solving CPP. The experimental results demonstrate the effectiveness of the proposed algorithm.

ABSTRACT. Remaining useful life(RUL) prediction of Lithium-ion batteries(LIBs) plays an important role in battery management system, and accurate prediction can ensure the safe and stable operation of battery. However, an accurate RUL prediction is difficult to achieve. In this paper, a method based on grey wolf optimization(GWO) and support vector regression(SVR) has proposed, which effectively improves the accuracy of LIBs remaining useful life prediction. Since the kernel parameter of SVR is difficult to select, GWO algorithm is employed to optimize SVR kernel parameters. This method is verified according to battery data set provided by the NASA Prognostics Center of Excellence(PCoE). Compared with SVR, the RUL prediction accuracy of GWO-SVR has been significantly improved. On this basis, compared with ALO-SVR, higher prediction accuracy is obtained. The accuracy of RUL prediction has been effectively improved.

ABSTRACT. In order to solve the management of the change back meter carrier module from the fault or renovation project, avoid serious waste of module resources, and cause a lot of economic loss, this paper designs an change back meter carrier module intelligent sorting and management techniques based on ARM. According to the actual operation test verification, the equipment has stable performance, simple operation and strong practicability, and fully meets the requirements for intelligent, automated and digital management of the change back meter carrier module; Through the power supply station, the change back meter carrier modules are sorted, the module sorting chain is shortened, the sustainable utilization rate of the equipment is improved, and the cost reduction and efficiency increase at the management level are realized.

ABSTRACT. With the rapid development of unmanned aerial vehicle (UAV) technology, the UAV surveillance system has attracted extensive attention in the intelligent transportation community. In this paper, an object detection YOLO model with global context cross (YOLO-GCC) is proposed for identifying small sized traffic elements in UAV image sequences. The concept of the asymmetric convolution is introduced to increase the robustness of the object detection model. Moreover, a global context attention module is added to extract more efficient features to ensure the real-time performance while improving the detection accuracy of small objects. The evaluation and comparison results on multiple UAV datasets demonstrate the effectiveness of the proposed model. Furthermore, an intelligent traffic signal scheduling algorithm named Traffic Deep Q-Network(Traffic-DQN) using deep reinforcement learning is introduced, which utilizes the traffic flow data obtained from YOLO-GCC as the benchmark for traffic scheduling. The experimental results demonstrate that the proposed algorithm can effectively alleviate traffic congestion compared with other methods.

ABSTRACT. The quality of Aluminum Profiles is the most important evaluation criterion in industrial production. To perform the quality control of Aluminum Profiles, strict defect detection must be carried out. Traditional machine learning methods need to design hand-crafted features in advance. Deep learning methods need to preset anchor parameters according to all defects, which is inefficient and inaccurate. In this paper, we propose an adaptive anchor network with an attention-based refinement mechanism for defect detection. The network has learnable parameters to generate anchors adaptively. Meanwhile, to better represent the different defects, we design a refinement module with the channel and spatial attention mechanism and deformable convolution at the stage of feature extraction. Besides, we also use cascade detection architecture to retain more defect information. The proposed method gets the AP of 62.4 and AP@50 of 86.1 on an industrial dataset, which has AP of 12.8 and AP@50 of 17.8 improved to the conventional methods and outperforms several state-of-the-art methods.

ABSTRACT. Interactive image segmentation can improve segmentation performance using manual intervention. Traditional interactive segmentation methods have unsatisfactory segmentation accuracy for images with complex background. Deep learning-based methods depend on large and accurate annotated datasets. In this paper, we propose an online interactive segmentation method based on graph convolutional network (GCN), which includes the superiorities of these two types of methods. We present a pre-segmentation stage to get an initial segmentation of the image, then propose an interactive GCN (iGCN) module to further improve the accuracy of the initial segmentation. Moreover, iGCN module is trained online without any pre-training burden. Experimental results show that our method outperforms several state-of-the-art methods on GrabCut and Berkeley datasets.

ABSTRACT. With the acceleration of the pace of modern life and work, people's psychological pressure generally increases, and the prevalence rate of depression is on the rise in recent years. It is becoming an area of research in the medical field to prevent and diagnose depression efficiently. In this paper, a depression analysis model based on BiAttention-GRU network is proposed. Firstly, the model obtains the key local features of text, sound and facial information through the parallel attention model, which reduces the influence of irrelevant information. Then, according to the contribution of these three types of features, the fusion feature vectors of the three types of features are obtained through the fusion attention model. Then inputing the fusion features of continuous frames to obtain the temporal information of the overall feature through the GRU network. Finally, the model calculated the depression prediction results through the Softmax layer. the results show that the model proposed in this paper is superior to the other three models by comparing this model with traditional LSTM network, CNN network and SVM, and the prediction accuracy is 89.77%.

ABSTRACT. In this paper, an intelligent image research method is introduced to conduct intelligent analysis on the imported shale gas reservoir images, distinguish organic pores and inorganic pores according to the degree of aggregation, and search the contour of the original image. The contour is fitted into a rectangle, and the two types of pores are respectively divided into circular pores or elliptical pores according to the ratio of length to width. At the same time, according to the pore area, the large pore, medium pore and small pore are distinguished and the corresponding number is counted. Finally, the analysis results are output. On the basis of the above pore identification and analysis, crack identification and analysis are carried out. According to the requirement of the ratio of length to width of cracks, cracks in the image are identified and relevant data are counted. At the same time, connectivity is analyzed to identify all connected pores. It realizes the fine qualitative and quantitative analysis of the type, size, number, spatial distribution and connectivity of pores and cracks in the clear image of shale gas reservoir.

ABSTRACT. We propose a specific lidar odometry system optimized with ground and road edges, GR-LO, for unmanned ground vehicle (UGV) navigation in the GPS coordinate system. Lots of lidar odometry (LO) and lidar-based simultaneous localization and mapping (SLAM) system have been proposed. Some of these methods can achieve high performance to estimate the lidar odometry. However, when we need to convert the pose of lidar from the lidar coordinate to the GPS coordinate system for navigation, there is usually a non-negligible error. There are usually two reasons for this error. The one is that the transformation between the two coordinate systems is not calibrated correctly. The second is that the cumulative error generated by the system. Our method can reduce the error caused by the first reason by extracting ground plane and road edge features in the specific environment and adding constraints to the pose graph. We have tested the proposed method with our datasets. The results show that the method can effectively reduce the absolute errors between lidar pose and GPS trajectory compared to several effective methods.

ABSTRACT. The Internet of things online monitoring technology has been widely used in high voltage transmission line temperature monitoring, environment, and the size of the amount of data. However, as the line is affected by the size of data and monitoring requirements, there is the possibility that multiple lines will have delay time at one time and a certain line will appear abnormal repeatedly. Manual processing speed is slow, low efficiency, and then bring huge losses. In order to improve the data transmission efficiency, make the line temperature prediction faster and grasp the line temperature more accurately, a temperature prediction method based on LSTM-ELM network under broad band and narrow band fusion is proposed to unify the broadband and narrowband structure and transmit data of different sizes through a unified frequency band. Establish the LSTM-ELM network, extract the characteristics of data in the time dimension, analyze the temperature data, and realize the rapid and accurate prediction of the line temperature trend. The experimental results show that the network prediction accuracy based on LSTM-ELM reaches 93.32% while the prediction time is greatly reduced to 875.75s, which can provide a reliable basis for background management in engineering practice.

ABSTRACT. Deep neural networks are frequently used to automate the examination of radiographic images in medical. These approaches may be used to train on huge datasets or extract features from small datasets using pre-trained networks. Due to the lack of large pulmonary tuberculosis datasets, it is possible to diagnose tuberculosis using pre-trained deep convolutional neural networks. Thus, this article aims to detect and diagnose tuberculosis in chest X-rays by combining a pre-trained deep convolutional neural network with a machine learning model. Combined the deep pre-trained DenseNet201 network with the machine learning XGBoost classifier to create a hybrid model for classifying patients as tuberculosis infected or not. The proposed model extracts feature using the pre-trained DenseNet201 neural networks and classify them employing the XGBoost classifier. We performed extensive experiments to assess the performance of the proposed DenseNet201-XGBoost model using tuberculosis chest X-ray images. Comparative study shows that the proposed DenseNet201-XGBoost-based tuberculosis classification model outperforms other competing approaches.