Search

Copyright of Coal Science & Technology (0253-2336) is the property of Coal Science & Technology and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

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

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

Copyright of Coal Science & Technology (0253-2336) is the property of Coal Science & Technology and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Copyright of Automobile Technology & Material is the property of Automobile Technology & Material Editorial Office 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.)

Early detection and distinguish of mine fires can avoid or reduce casualties, property damage, and secondary accidents. There are no natural light sources such as sunlight, moonlight, starlight, and lightning underground, and the main factor affecting the recognition of mine fire images is the mine light source. Circularity can eliminate interference from circular light sources, but it is difficult to exclude interference from non-circular light sources. Rectangularity can eliminate interference from rectangular light sources, but it is difficult to exclude interference from non-rectangular light sources. In engineering practice, due to the different shooting angles of the camera, the image of the mine light source may deform and cannot present an ideal regular shape. It is difficult to eliminate the interference of the mine light source using circularity and rectangularity algorithms. It has been revealed that the area of the external connection graphic area of the flame image is significantly larger than the actual area of the image, and the external connection graphic area of the actual light source image of the mine, such as circular lamps, rectangular lamps, and square lamps, is approximately equal to the actual area of the light source image. In this study, a mine fire recognition and anti-interference method is proposed based on the internal concavity of image, the ratio of the target image area to the external connection graphic area of the image (i.e. internal concavity of image) is calculated, and the flames and mine light sources are distinguished based on the small concavity value in the flame image and the large concavity value in the mine light source image. The internal concavity method proposed in this paper is not affected by the distance from the camera to the detection target and the size of the image, the installation position and angle of the camera to capture the detection target, and the shape of the mine light source. It has strong adaptability and high recognition accuracy. The experiment shows that the internal concave recognition method calculates the maximum average difference between the mine interference light source and the flame image, with the smallest fluctuation and the best discrimination. It is also least affected by the camera’s shooting angle and distance, and has the strongest anti-interference ability, with an accuracy of 91.6%. The rectangular recognition method calculates the average difference between the mine interference light source and the flame image, with small fluctuations and good discrimination. It is less affected by the camera’s shooting angle and distance, and has average anti-interference ability, with an accuracy rate of 72.5%. The roundness recognition method calculates the minimum average difference between the mine interference light source and the flame image, with the highest fluctuation and the worst discrimination. It is less affected by the camera shooting distance and more affected by the camera shooting angle, and has the worst anti-interference ability, with an accuracy of 12.0%. Therefore, the internal concave recognition method proposed in this paper is superior to rectangular and circular degrees, with the best discrimination, minimal influence from camera shooting angle and distance, and the strongest anti-interference ability.

The accurate identification and positioning of the anchor position of the coal mine tunneling roadway is the key technology that the drilling anchor robot urgently needs to break through to achieve intelligent permanent support. An intelligent identification and positioning method for roadway anchor hole position based on visual image and laser point cloud fusion is proposed, which includes three steps: image target recognition, point cloud image feature fusion, and positioning coordinate extraction: ① In order to solve the problem of blurred image of anchor hole contour caused by environmental factors such as low illumination, water mist and dust in coal mines, the IA (Image-Adaptive)-SimAM-YOLOv7-tiny network was used to visually identify the position of the anchor hole to be anchored in the roadway, which can adaptively enhance the image brightness and contrast, recover the high-frequency information at the edge of the anchor hole, and make the model focus on the characteristics of the anchor hole, so as to improve the success rate of anchor hole detection; ② The Region of Interest (ROI) of image detection is generated by perspective projection relationship to generate a cone-shaped area of interest to obtain the corresponding target point cloud cluster; ③ The point cloud processing algorithm is used to extract the anchor hole boundary point cloud, obtain the central coordinates of the hole position and its normal vector, and judge the correctness of the anchor hole recognition by comparing the coordinate depth difference. In this paper, a drilling and positioning system for bolting trolley manipulator is built to verify the accuracy and accuracy of the algorithm's autonomous positioning, and the experimental results show that the mean average precision (mAP) of the IA-SimAM-YOLOv7-tiny model is 87.3%, which is 4.6% higher than that of the YOLOv7-tiny model. Compared with the single vision method, the fusion algorithm proposed in this paper has a positioning error of 3 mm, and the average recognition time of the system in the case of a single anchor hole is 0.77 s, compared with the single vision method, the fusion of laser and visual multi-source can not only reduce the influence of environment and small sample training on positioning performance, but also obtain the normal vector of the anchor hole position, which provides a basis for the manipulator to adjust the drilling posture to achieve accurate anchoring.

Copyright of Food & Machinery is the property of Food & Machinery Editorial Office 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.)

With the leapfrog development of the railway industry in China, the operating speed and density of the railways increase gradually, and large tamping machinery for track maintenance is used more widely. Traditional manual tamping truck has high requirements for operators and high working intensity. In order to achieve the goal of reducing the intensity of manual work and ensuring the efficiency of on-site construction, this paper proposes a design scheme of intelligent automatic tamping system based on DCL-32 tamping truck without changing the function and mechanical structure of the original machine, and focuses on the analysis of the railway track sleeper detection and positioning of the pick. In this scheme, the YOLOX model is used for image detection, and the recognized results are checked with the output mileage of laser ranging sensor, to complete the detection of track sleepers quickly and accurately. Based on the historical data of tamping truck braking distance during railway maintenance, this paper fit a braking relationship curve to predict the tamping position. The results of field prototype loading test show that the speed of automatic tamping is equal to that of manual tamping under single and double tamping modes. The control scheme of automatic tamping system proposed in this paper has a good trial effect, which lays a foundation for subsequent product promotion.

[Purpose/Significance] We used to process soybean leaf data by looking at them and process data manually, but this method is very inefficient. In order to improve the classification accuracy and efficiency of soybean leaf images, further for storage and management of these images, we used the deep learning technique to make an in-depth study of text data and image data of soybean leaves for the image recognition and classification. The application of deep learning in agricultural data management mainly focuses on the image recognition and classification of plants and plant phenotypes in large-scale data, detection and classification of agricultural diseases and pests, detection and classification of crops and weeds, and prediction of crop yield. Through case analysis, our sample data demonstrated the application process of deep learning technology. [Method/Process] This paper systematically described the whole process of classification and recognition of agricultural data by using the deep learning technique. Through recognition and disease monitoring of plant leaves, the leaf morphology of soybean plants in the soybean experimental field of Heilongjiang Academy of Agricultural Sciences was taken as an example. We analyzed the image features of soybean leaf morphology, and carried out the classification and recognition research of soybean leaf morphology based on deep learning. Deep learning techniques have replaced shallow classifiers that use manual feature training and can identify soybean leaves with a high degree of accuracy as long as sufficient data are available for training. We adopted DenseNet model, which is suitable for common network model. The advantages of this model are that it has the best performance and the least storage requirements. First,we selected support vector machine (SVM) and random forest (RF) in traditional machine learning methods to identify soybean leaf morphology. Second, AlexNet and ResNet were selected to identify soybean leaf morphology. Finally, the recognition accuracy of different methods were compared with the DenseNet network adopted in this paper. [Results/Conclusions] Through the training of DenseNet model, the recognition accuracy of 94% was achieved, which successfully solved the problems of long time, low efficiency and low recognition accuracy of traditional methods in processing image classification of soybean leaves, and could meet the actual needs of agricultural image data classification. Future research efforts will strive to collect a wide range of large and diverse data sets to facilitate soybean leaf recognition, and focus on developing reliable background removal techniques and incorporating other forms of data to improve the accuracy and reliability of soybean leaf recognition systems.

Copyright of Journal of Frontiers of Computer Science & Technology is the property of Beijing Journal of Computer Engineering & Applications Journal Co Ltd. 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.)

Under the background of intelligent construction of coal mine, this paper studies and discusses the construction of AI platform architecture system of the whole coal mine based on video perception, network and information technology as the media, artificial intelligence and big data as the technical support; according to the actual situation of coal mine, the real application scenario is mined and analyzed; the paper carries on research on intelligent sensing, deep learning and intelligent analysis for coal mine equipment status, personnel behavior and environmental changes to realize intelligent alarm and control; the paper researches and analyzes relevant key technologies such as artificial intelligence and deep learning.

Aiming at low accuracy of recognition in complex leakage images of subway tunnels, this paper studied the application of image recognition algorithms in tunnel water-leakage disease, was using two kinds of neural networks: U-Net and PSP-Net. Based on the subway tunnel images, tunnel leakage disease sample data set was constructed, where 5000 subway tunnel disease image data and 4000 labeled images were included. This data set could be used for deep neural convolution network training. U-Net and PSP-Net deep learning frameworks were established, and the neural networks were trained by self-made data set. The accuracy of the prediction results from U-Net and PSP-Net was evaluated, and the performance of the two networks in the application of leakage disease identification was compared. U-Net had stronger coverage for the disease and quick detection, while PSP-net had higher accuracy fit to fine detection. [ABSTRACT FROM AUTHOR]

Copyright of Electric Power Engineering Technology is the property of Editorial Department of Electric Power Engineering Technology 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.)

Grain surface movement monitoring is an important part of the daily inventory inspection of national grain reserves, and is a new requirement for grain storage supervision. In order to solve the technical problems of the traditional grain surface movement monitoring, this paper proposed a laser radar-based grain surface movement monitoring method. A full-size experimental platform that can simulate the real grain pile state was established, and a high-precision laser three-dimensional measurement device was utilized to design an algorithm to identify the abnormal movement of the grain surface of the bulk grain pile based on the laser scanning point cloud data and the information of entering and exiting the warehouse operation. The algorithm was examined and verified through the experimental platform. The results showed that: the method can directly obtain high-precision coordinate information of the grain pile surface, which overcome the problem of insufficient data accuracy of image recognition technology. The constructed algorithm of grain surface motion discrimination based on the coordinate information of the grain pile shape and the data of entry and exit operation status was capable of realizing reliable quantitative computatio. The proposed method was applied to the actual grain warehouse scenario, which verified the feasibility and validity of the method, and it was able to satisfy the dynamic grain inventory supervision needs. The proposed method was applied to the actual grain silo scenario to verify the feasibility and effectiveness of the method, which can meet the demand of grain stock dynamics supervision and provide a new technology for online monitoring and early warning of grain surface variation.

Copyright of Coal Science & Technology (0253-2336) is the property of Coal Science & Technology 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 computer technology related to unmanned vehicle driving is developing rapidly, and the vehicle identification technology is more and more widely used in intelligent traffic monitoring and intelligent traffic information systems. Vehicle identification includes sports vehicle detection, type classification, vehicle tracking and so on. In this paper, the key technologies such as vehicle identification detection in smart car unmanned technology are deeply explored. Using the image recognition result of the vehicle, this paper chooses the CAM offset algorithm based on the average color difference value as the basis for identification. The algorithm uses the vehicle color as the feature to perform the tracking detection method, and does not follow the light and dark of the vehicle. The shape change is disturbed, and the reliability in the vehicle identification process is high. This algorithm is improved to achieve better recognition of the vehicle.

Based on the TensorFlow framework, this paper builds convolutional neural networks to recognize the defects in the electroluminescent images of cells. It selects the exposed data set that contains the different types of defects in the cell. Based on the traditional VGGNet network, the full convolution neural network is used for training, and this paper analyzes the training effects of different loss functions and dropout probabilities on data set. Experiments have shown that the algorithm accurately recognizes whether the cell is defective. The study also shows that the compression network structure greatly increases the training rate of the algorithm, which makes the simplified model more portable and provides an effective solution for a wide range of real-time defect recognition.

Due to the various species of fish, and the influence of different light and background environments, the recognition accuracy of some traditional fish recognition algorithms bases on color texture or feature point extrac tion is reduced, and good classification results cannot be achieved. To solve the problem, based on the existing Alex Net convolutional neural network, this paper proposes one method to speed up model training by reducing redundant convolutional layers. An item-based soft attention algorithm is applied to the improved AlexNet convolutional neural network model, which consists of four convolutional layers, one item-based soft attention layer, and two fully-con nected layers. Meanwhile, a fish recognition model is established using transfer learning methods. The test results show that the average recognition accuracy of the proposed algorithm achieves 97. 43%, which is 4.08% higher than the original AlexNet model, and the average recognition rate of some fishes achieves 99.31% and time consumption of fish recognition is reduced by 35%. In conclusion, compared with state-of-the-art fish recognition algorithms. the fish recognition algorithm proposed in this study achieve higher accuracy, lower model complexity, and stronger robustness. [ABSTRACT FROM AUTHOR]

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

Copyright of Journal of Harbin University of Science & Technology is the property of Journal of Harbin University of Science & Technology 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.)

Similar material model experiment is an important means to study mining subsidence and damage, and it is very important to obtain high-precision model deformation information in the experiment. There are still some limitations in the accuracy and efficiency of data acquisition in many existing model deformation measurement methods. In this paper, the simulation experiment of similar materials is carried out by taking the circular mining and filling mining of multiple coal seams in a mine in northern Shaanxi as models, and the high-resolution sequence images are obtained by using the non-measurement camera at close range. Through image distortion correction, automatic stitching and feature recognition, the deformation of similar material models can be obtained with high precision. Firstly, the influence of model measuring point marks and camera parameters on deformation measurement accuracy is analyzed, and the best model shooting scheme is determined, that is, the non-measurement camera is used to take an orthographic photo at a distance of 1.5−2.0 m from the model, forming an image sequence with a horizontal overlap of more than 55% and a vertical overlap of more than 30%; Furthermore, the bilinear interpolation method is used to correct the distortion of the obtained image, the automatic and robust image mosaic algorithm is used, the Harris corner detection operator is used to extract the feature points, the RANSAC algorithm is used to classify the pseudo-matching points, and a series of automatic processing processes such as color transition and smoothing at the butt edges are used to mosaic the image sequence. The resolution of the generated panoramic image of the model is about 10 times higher than that of the conventional fixed camera. In the process of image processing, the measuring point mark is used as the training network, the confidence level is added based on Faster R-CNN algorithm, and the appropriate sliding template is selected for region analysis, which significantly improves the recognition rate and accuracy of the measuring point mark. The results show that the precision of measuring point coordinate extraction in model experiment is better than 0.027 mm, which is equivalent to the field measurement precision of 5.4 mm, which fully meets the requirements of mine similar material model experiment. By making an experimental model with a scale of 1∶200, the deformation and failure characteristics of overlying strata caused by cyclic mining of multiple coal seams are simulated and analyzed, and the maximum surface subsidence reaches −18.937 mm; When the mined-out area is filled, the lower part of overlying rock is slightly deformed, and the surface subsidence is not more than 2 mm. The experimental results provide an effective technical means for the efficient and accurate collection of experimental data of similar material models in mines.

Objective： In order to meet the needs of the edge computing of the Internet of Things, a recurrent neural network algorithm was introduced for the first time in this paper, and an intelligent real-time classification and recognition system was constructed to conduct research on food packaging images. Methods： To build a simulation experiment test model, firstly preprocessed the image data set, de-redundantize, grayscale, and normalized the two-dimensional image, and finally input the time-sequential data in parallel. Using a typical memristor as the research object of realizing hardware RNN, the nonlinear function of memristor was used to construct the mapping layer of parallel array reserve pool neural network. the ridge regression algorithm was used to solve the problems of overfitting in the training process. Results: The classification accuracy of the food packaging data set was as high as 98.59%. Conclusion： The system reduces the number of traditional neural network layers, reduces the training cost, and realizes high-precision real-time online recognition of time series signals.

In order to effectively solve the problems of independent construction of various systems in the existing coal industry, lack of sharing of multi-system data, low utilization of intelligent analysis and data mining, this paper uses the Internet of Things, big data, artificial intelligence and other information technologies, and on the basis of in-depth integration of coal mine safety management business, we proposed the construction scheme of the whole process traceable coal mine safety intelligent management and control cloud platform, focusing on the analysis of risk management and control, hidden danger trouble shooting, team task processing business processes such as violation correction. The platform realizes the storage, integration and analysis of multi-source heterogeneous monitoring data through the data center, integrates the nine subsystems involved in coal mine safety management into the safety intelligent management and control platform, and uses the data to share a chart for collaborative management and control and effective analysis and display of data. The application results show that the platform has realized coal mine data fusion, collaborative control, business linkage and decision-making, and has a good effect in improving safety control efficiency, reducing control costs and maximizing enterprise interests.

For the effective, convenient and unified management of materials about military logistics, this paper studies and designs a monitoring system for monitoring the presence of items. The system uses the motherboard of Raspberry Pi to collect the level signal of the infrared sensor about the presence of the items and the images taken by camera of the items. Then it converts the data into the general data frame, and sends the frame to the data processing module through the source-specific multicast. Finally, the image recognition based on the convolutional neural network is used to judge the correctness of the item, and display its status in real time through the monitoring interface. It has been verified that the system can ensure the real time of data acquisition and the accuracy of identifying items. It has strong practicability.

Relay protection device is an important part to ensure the safe and stable operation of power system. With the rapid increase of the number of substations and relay protection devices,the daily inspection workload of the operation and maintenance personnel has become saturated,which can not guarantee the high quality and no dead angle inspection every time and brings hidden dangers to the reliable operation of the protection devices. In this paper,an intelligent inspection technology of protection device based on convolution neural network image recognition algorithm is proposed. With the help of the cameras installed in the front and back of the cabinet,the unmanned or few people inspection of the protection device can be realized. Firstly,the intelligent inspection system of the protection device is introduced,and the intelligent inspection items that can be realized is analyzed. The conclusion that convolution neural network can be used for image recognition is drawn. Secondly,taking the platen state recognition as an example,the training sample set and test sample set required by the inspection items are introduced,and the convolution neural network level of the inspection items is given. Then the training sample set is used to train the convolution neural network of different inspection items,and finally each network is tested. The test results show that the neural network image recognition rate of each inspection item is above 96%,even 98%,and the recognition effect is good.

As the traditional power grid has evolved into smart grid， testing of relay protection devices has shifted from manual mode to automatic model. By analyzing the shortcomings of the existing automatic testing systems， the paper proposes a new relay protection testing system based on 5G and image recognition technology to reduce labor cost and improve the testing efficiency of relay protection. The new system adopts a modular design idea and uses 5G network to connect the cloud server， automatic testing platform and mobile phone terminal so that testing solutions can be downloaded on the client， remote and real-time testing monitoring can be achieved， and testing report can be generated fast. In order to quickly obtain the pressure plate status of the protection device， image recognition technology is introduced， and corresponding solutions are proposed for the intelligent generation of test solutions， automatic configuration of virtual terminals and other key technologies. The system has been successfully applied to the acceptance of 500 kV Xiaowei substation and 220 kV Lyuting substation with satisfactory test results.

Aiming at the poor measurement effect of behavior monitoring in real life,a new method of extracting human behavior features is proposed.It not only considers the body point information,but also integrates the environmental attribute information of image.Considering that a large number of existing experiments use a variety of complex algorithms for experimental classification on the basis of human body features extraction,it does not take into account the irrationality of only using the body features for algorithm evaluation.Therefore,an image information reconstruction method based on body features is proposed in this paper,which combines the image convolution network of body features,attention mechanism and image recognition method to realize human behavior recognition.The body point information is extracted by Openpose,and then the body points are classified by graph convolution and attention.On the basis of the first classification,the body point expansion coefficient is added to segment the images so as to realize second accurate classification.Finally,the evaluation accuracy on the HMDB51 dataset improves by 5.6%,and it has a big advantage in the actual test.This shows that the method is not only more accurate,but also has more practical application value.

To solve the problem of slow response and low accuracy in the traditional image recognition method of transmission line target inspection,an improved faster-region convolutional neural network (Faster-RCNN) deep learning recognition algorithm is proposed. In this paper,the image features are extracted by zeiler and fergus net (ZFnet) and the ZFnet model parameters are reset to obtain more target details. Then,the Faster-RCNN is used to detect the target. The target candidate box is generated by the sub-network region proposal model and the parameters are tuned by the fast-region convolutional neural network (Fast-RCNN). In addition,the refining stage is introduced into the output part of the Faster-RCNN to increase the refinement of classification and regression of the target features. And then the multiple bounding boxes with the target are combined to achieve accurate classification and coordinate positioning. The results of the experiments show that the improved Faster-RCNN algorithm can effectively identify the transmission line equipment and its defects. The overall recognition rate of the method could reach 93.5% within 1 s of the response time. Compared with the image recognition and the deep learning such as single shot multibox detector (SSD) and you only look once (YOLO),the proposed algorithm improves the identification accuracy and response speed of power equipment,and has certain advantages in intelligent inspection of transmission lines.

Taking the gas drainage system as the platform and the industrial pointer instrument as the design object, this paper proposes and designs an automatic identification device for industrial instruments, which uses the camera to obtain the instrument image, obtains the instrument reading and local display through image recognition technology, and has CAN bus transmission. It solves the problems that a large number of industrial instruments in the gas drainage field need manual inspection, and do not have the data communication ability, and cannot communicate with the existing gas drainage system, and the purpose of accelerating the construction of intelligent mine has been achieved.

To solve the complexity of convolutional neural network and the large number of parameters in image recognition task, this paper proposes a lightweight network SepNet. In this structure, the traditional fully-connected layer is replaced by Kronecker product in the classifier module. In order to further optimize network structure, in the feature extraction module, by balancing the depth and width of the network, a separable residual network module using the deep separable convolution and residual network is designed. Finally, a lightweight network architecture which can realize end-to-end training is formed, which is called sep_res18_s3. The experiments are conducted on MNIST, CIFAR-10 and CIFAR-100 datasets respectively. The results show that compared with the VGG10 network, the designed SepNet can reduce the number of parameters and computation by 94.15% without losing its accuracy. At the same time, compared with cov_res18_s3, sep_res18_s3 can still reduce the parameter amount by 58.33% and 81.82% of FLOPs. Experimental results show that replacing the fully-connected layer with Kronecker product can not only maintain the accuracy of training results, but also significantly reduce the number of parameters and calculation costs, and to a certain extent, it can prevent overfitting. On this basis, combining the deep separable convolution and residual structure, it proves the effectiveness of sep_res18_s3.

In order to solve the problem that visible light images are greatly interfered by mine lights, car lights, roadway lights, dust and red object in mines, the mine external fire sensing method based on near-infrared and visible light images is proposed. The near-infrared images and visible light images are fused to bring into play the advantages of obvious early flame characteristics in near-infrared images and obvious mid-term flame characteristics in visible light images so as to improve the recognition rate of mine external fire. This paper proposes a mine external fire sensing method based on multiple information fusion of near-infrared image, visible light image, temperature sensor, gas sensor and smoke sensor. The method fuses near-infrared image, visible image and sensor information. The sensor does not need no blind area coverage. Only a certain number of gas sensors, temperature sensors and smoke sensors need to be set up, which has the advantages of being timely, reliable and low-cost.The fire alarms are divided into four categories, including the image, temperature, smoke and gas fire alarms. The image fire alarm information includes near-infrared image and visible light image information. When the fire is detected by near-infrared image or visible light image, the image fire alarm signal is issued.When the temperature exceeds the limit, the temperature alarm signal is issued. When the smoke exceeds the limit, the smoke alarm signal is issued. The gas fire alarm information includes CO, CO2, O2, etc. When CO, CO2 or O2 exceeds the limit, the gas alarm signal is issued. The fire alarms are divided into four levels, including blue, yellow, orange and red, with red being the highest level.When one of the four alarm signals of image, temperature, smoke and gas issued, the blue alarm signal is issued. When two of the four alarm signals of image, temperature, smoke and gas issued, the yellow alarm signal is issued. When three of the four alarm signals of image, temperature, smoke and gas issued,the orange alarm signal is issued. When all four alarm signals of image, temperature, smoke and gas issued, the red alarm signal is issued.

Abstract： Deep learning and neural network models have been new research directions and hot issues in the fields of machine learning and artificial intelligence in recent years. Deep learning has made a breakthrough in the applications of image and speech recognitions, and also has been extensively used in the fields of face recognition and information retrieval because of its special superiority. Bone X-ray images express different variations in black-white-gray gradations, which have image features of black and white contrasts and level differences. Based on these advantages of deep learning in image recognition, we combine it with the research of bone age assessment to provide basic datum for constructing a forensic automatic system of bone age assessment. This paper reviews the basic concept and network architectures of deep learning, and describes its recent research progress on image recognition in different research fields at home and abroad, and explores its advantages and application prospects in bone age assessment.