Research On Substation Violation Recognition Based On Video Structure

Video structure is one of the focuses and hots-pots in the field of computer vision and security.It can quickly extract target information from massive videos and is widely used in practical tasks.At present,video structured technology has emerged in the field of safety production in the power industry.It effectively solves the problem of manual detection of substation operation violations,and further accelerates the power industry to an intelligent era.Therefore,this thesis aims to use the basic theory of object recognition method based on video structure to build a fast and efficient recognition network,in order to realize the efficient recognition of multiple conventional violations in substation scene,and then ensure the safety of substation operation.Due to the wide field of vision of substation scene,the target volume to be detected is small and the scale is uneven.At the same time,according to the characteristics of its business needs,the detection system is required to have certain real-time performance.To solve these problems,based on the lightweight target detection model,this thesis focuses on the target detection problem of uneven scale distribution in large visual field of substation scene.On this basis,the proposed model is further compressed to meet the needs of real-time identification in actual business.The main work of this thesis is as follows :(1)An adaptive anchor optimization method based on improved differential evolution is proposed to solve the problem of inaccurate anchor frame positioning caused by uneven distribution of data,which leads to the decrease of target recognition accuracy.In order to better capture the relationship between the proportion and number of target objects,the ratio of width to height of the boundary box is selected as a variable,and the sum of the nearest distance between Anchor and the real boundary box is calculated as an adaptive function.In addition,in order to avoid the category of Anchor orientation with a real box sample,the weight value is added to the distance calculation.The effectiveness of this method is verified by the comparison experiment of loss function and visual scatter plot.(2)A target detection method based on multi-scale attention is proposed to solve the problem of network false detection and missed detection caused by the large field of vision of substation scene or the small size of the target itself.The original network is a shallow network structure,which is very unfriendly to small targets.Adding a detection layer in the network can effectively reduce the area of the network output feature mapping on the input image,so as to enhance the ability of the network to extract the target feature and improve the probability of target detection.Attention mechanism is introduced to enhance the networks attention to small targets and reduce the false detection of targets.By comparing the effects of different attention models and their positions in the network on the overall performance of the network,the attention model with the best effect is selected to be added to the network.Finally,the two parts of the improved content are fused,and the normalized weight idea is used to moderately adjust the relevant parameters of the attention model,so that the overall network achieves the optimal recognition effect.(3)Structured pruning method is used to compress the model proposed in content 2 to meet the needs of real-time detection in substation scenarios.Through in-depth study of the characteristics of model compression method,the structural pruning method is used to prune the complex network model to eliminate redundant channels and reduce the overall computational complexity of the network.While ensuring the accuracy of model identification,the running speed of the model is improved,and the model has the ability of real-time detection.In summary,in order to achieve efficient detection of multiple violations in substation scenarios,this thesis proposes an adaptive anchor frame optimization method and a target detection method based on multi-scale attention.In addition,in order to meet its business needs,the structured pruning ensures the accuracy of network identification and the ability of real-time detection.