Research On Video Cleaning Algorithm In Coal Mine Underground Based On Deep Learning

Safety production of coal mine and disaster prevention are always been important issues emphasized in our country,intelligent video surveillance and monitoring(VSAM)technology is one of the most effective ways to prevent and relieve this problem.The continuous increase of underground video surveillance equipment in coal mines has led to the production of a large number of near-duplicate video data.The quality of video data has become increasingly prominent,and the maintenance and management of video data has become more and more challenging.In this project,we focus on the issue of video data quality of near-duplicate video in coal mine,they are studied by using the video data set of coal mine.The concrete contents include:(1)Although the existing near-duplicate video retrieval methods can effectively identify near-duplicate videos,it is difficult to automatically clean the near-duplicate video data under the premise of ensuring data integrity in order to improve the quality of the video data.This paper proposes a near-duplicate video cleaning method combining VGG-16 network and FD-Means clustering.This method uses the VGG-16 network model to extract High-level Semantic Feature of the video;in the unsupervised video cleaning task,because the K-Means algorithm need to predetermine the number of clusters K,and the video content is random,the number of clusters cannot be determined.This paper proposes a FD-Means clustering algorithm model.When updating clusters,according to the distance between the data point and the cluster center,the discrete points with a distance greater than the distance threshold are consttructed as a new cluster;in order to eliminate clusters that are too close,the distance between each cluster center is smaller Finally,the near-duplicate video data outside the cluster center point is cleared.Experimental results show that this method can effectively solve the problem of near-duplicate video data cleaning,improve the data quality of the video,and reduce the waste of storage resources.(2)Because it is difficult to guarantee the quality of underground videos in coal mine,and it is difficult to accurately represent the high-level semantic features of videos using only the video spatial features extracted by the VGG deep network model.To this end this paper proposes a multi-attention residuals network,which combines spatio-temporal features and attention models to improve the representation ability of video features.This method first selects Resnet34 embedded in CBAM as the basic network to improve the feature extraction of the salient areas of the image;secondly,it uses a long-term short-term memory network with time attention to capture the timing characteristics of the video frame sequence;finally,two networks are connected in series to construct a multi-head attention Resnet.This method has been experimentally verified on CC_WEB_VIDEO and coal mine datasets.The experimental results show that the multi-head attention Resnet proposed in this chapter has increased F1-score by 18%and 7%in comparison with the comparative experiments in the tasks of CC_WEB_VIDEO and the near-duplicate video detection on the coal mine dataset.In the video cleaning tasks on CC_WEB_VIDEO and coal mine datasets,compared with the pre-trained VGG 16 feature extraction method in Chapter 3,F1-score increased by 8.7%and 15.7%respectively,greatly improving the accuracy of near-duplicate video cleaning.