Research On Video Scene Classification Method For Coal Mine Based On Deep Neural Network

With the improvement of coal mine intelligence,the intelligent monitoring system is playing an increasingly important role in coal mine safety and production.Classify the coal mine video scenes can not only provide valuable reference information for coal mine video surveillance work,but also provide a basis for the investigation of coal mine abnormal events.Therefore,how to classify coal mine video scenes accurately has becoming one of the research hotspots in the field of coal mine video processing.Existing scene classification methods are rarely researched for coal mine video information.In this paper,coal mine video data with complex background was considered as the research object,combined with video feature extraction and self attention theory,coal mine video scene classification method based on deep neural network was researched.(1)In order to guide more discriminative features to participate classification,a video scene classification method with improved loss function was proposed.Firstly,the deep convolutional neural network was used for image denoising.Then,the CNN network and the bidirectional LSTM network were combined to learn the spatio-temporal context features of the coal mine video that were denoised by the deep convolutional neural network.Finally,the Center loss and Softmax loss were fused to achieve jointly supervise classification,and the Center loss was adopted as a new discriminative regularization term to retain the geometric distribution of intra-class and inter-class samples to achieve video scene classification of coal mine.(2)In order to obtain more scene information,a video scene classification method based on hierarchical self attention of coal mine was proposed.Firstly,the CNN network was used to extract the spatial features of the video,and through adding a self attention on spatial features to calculate the pixel relationship of the video image,so that the significant pixels of image can get more weight.Then the bidirectional LSTM network was used to extract the temporal features of the video,and the self attention was used one more time to calculate the weight of the video frame sequence feature weight so that the significant frame features in the video frame sequence can get more weight.Finally,the softmax loss and center loss were fused to output the multi-classification results of coal mine video scenes.(3)Finally,the coal mine monitoring video dataset was used to simulate and verify the proposed classification models.Experimental results showed that compared with different video scene classification methods,the classification method based on improved loss function and the hierarchical self attention can effectively improve the accuracy of coal mine video scene classification.At the same time,the center loss and attention were visualized,and its impact on video scene classification of coal mine was also analyzed.