Research On Improved Method Of Yolov5 Target Detection Algorithm In Complex Scene

Object detection is an important issue in the field of computer vision.It is more and more widely used in real-world scenarios.The continuous expansion of application requirements has prompted object detection methods to continuously develop in the direction of high performance,real-time,and diversification.Among them,YOLOv5 is a target detection technology that is widely used in medical imaging,unmanned driving,intelligent security and other fields.However,there are many interference factors in complex and changeable real-world scenes,and the detection effect of YOLOv5 is greatly affected.According to actual reality application problem,in this paper,the target scale on a wide range of change,background information and output the deviation of box three aspects YOLOv5 algorithm to improve the method of study,the main research content is as follows:(1)For the problem of large-scale scale changes of objects in complex scenes,carry out research on the improvement of YOLOv5 target detection algorithm based on multi-scale feature fusion.YOLOv5 fuses features through a path aggregation network in the multi-scale feature fusion stage,ignoring the differences between features at different scales.Therefore,an adaptive multi-scale feature fusion method is proposed to add the complete feature information before dimensionality reduction by adding skip connections,and add learnable adaptive weights to the input features of different scales to distinguish their contributions to achieve the purpose of improving the efficiency of multi-scale feature fusion,and then use the Pascal VOC standard dataset for comparison and verification.(2)For the problem of background information interference of objects in complex scenes,carry out research on the improvement of YOLOv5 target detection algorithm with joint attention mechanism.Through the research and analysis of several commonly used attention models,an adaptive pooling attention module is proposed for the traditional attention module that ignores the difference between average pooling and maximum pooling.Add an adaptive pooling attention module to the Neck part of YOLOv5,so that the model pays attention to the information difference of the feature map,and generates adaptive channel and spatial attention weights according to the importance of the information,so as to improve the network anti-interference information,The purpose of the ability to grasp the main features,and then use the Pascal VOC standard data set for control verification.(3)For the problem of offset of YOLOv5 output prediction frame in complex scenes,research on the improvement of YOLOv5 target detection algorithm based on weighted fusion of regression frames is carried out.YOLOv5 uses non-maximum suppression to screen the optimal box with the highest score in the post-processing stage of the regression box,but the screening result of the optimal box is not completely accurate.Therefore,it is proposed to use the regression box weighting method to use all the prediction box information to correct the final box.The output box position is used to improve the accuracy of the predicted box,and then the Pascal VOC standard data set is used for comparison and verification.Finally,these three improved methods are integrated into the YOLOv5 model,and a control experiment is carried out using the flame and smoke data set mixed in a real industrial project in an off-campus internship of a company in Nanjing to verify that the target scale in the actual scene changes greatly,the background information is cluttered and Target detection methods under disturbances such as output box offsets improve the effectiveness of the method.