Research On Camouflage Target Detection In Mine Background

Camouflage target,that is,the foreground target and the background in terms of color,texture and other characteristics,exist in a variety of different scenes,such as animal camouflage in nature,camouflage camouflage on the battlefield,overalls camouflage in the mine,etc.Due to the complex environment of the mine and the weak safety awareness of mine staff,safety accidents have occurred frequently in recent years,and the life safety of staff cannot be guaranteed.In view of this problem in the current mine farm,based on the YOLOv4 model,this paper constructs a camouflage target detection algorithm for mine site from four aspects: receptive field module,attention mechanism,feature fusion module and loss function.By detecting whether there is a camouflaged target in the danger area,the function of target cross-border warning is realized,so as to ensure the life safety of mine personnel.In addition,in order to facilitate the smooth landing of the research algorithm,the camouflage target detection system in the mine site is designed and realized,and the analysis results are visualized,which greatly facilitates the use of camouflage target algorithms by non-technical personnel.The specific research content of this paper is as follows:Firstly,in view of the imbalance of positive and negative samples in the target dataset of camouflage of the mine and the high integration of the work clothes worn by the mine staff and the background of the mine,this paper designs a multi-branch mixed convolution and improves the loss function.The multi-branch mixed convolution consists of 4branches,and the expansion convolution with different expansion rates and convolution kernels is set on each branch,and the feature maps output by the branches are aggregated by adding and Concat to obtain multi-scale large receptive field features.The confidence loss function and the regression loss function use different modulation coefficients to adjust the proportion of positive and negative samples and take into account the orientation matching problem of the prediction box and the regression box,which accelerates the convergence speed of the model.Secondly,aiming at the problem that YOLOv4 is difficult to use the mine to camouflage the effective features of the target,this paper improves the feature fusion structure and integrates the attention mechanism into YOLOv4.The characteristics of camouflage target determine its high degree of fusion with background objects,and feature maps with strong semantic information can better grasp the global features,this paper improves the fusion method between feature maps,realizes the adaptive fusion of deep feature maps and shallow feature maps by introducing spatial attention mechanism,improves the proportion of strong semantic information in the feature fusion process,and adds attention mechanism modules at the detection head,thereby improving the expression ability of features.Finally,the camouflage target detection algorithm proposed in this paper is experimented,and its effectiveness is verified by comparing with the existing algorithms from the qualitative and quantitative perspectives.In addition,in order to facilitate the smooth implementation of the algorithm in this paper,a camouflage target detection system is designed and implemented,which detects the video stream provided by Hikvision in real time and visualizes the detection results.