Improved Object Detcetion Algorithm Based On YOLOv5

With the proliferation of the Internet and the widespread use of electronic devices such as smartphones,computers and tablets in work and life,a wide variety of image information is generated.As a result,target detection has become widely used and an important research direction in image information processing.However,the presence of different types of targets to be detected in complex scenes,and the uneven distribution of the size and position of these targets,as well as the dense presence of overlapping targets,make detection more difficult and place higher demands on the ability of target detection algorithms to make better use of image information.To address the problem that YOLOv5’s backbone network pays undifferentiated attention to image information,causing the algorithm to focus on too much invalid information,an improved algorithm based on the coordinate attention mechanism is proposed.By inserting the coordinate attention module into the backbone network,the target detection algorithm is able to improve the extraction of valid information from the feature image types and reduce the interference of invalid information,so that the algorithm can accurately locate the detected objects while controlling the computational effort.Considering that the insertion of attention mechanisms at different positions of YOLOv5 s may have different effects,three arrangements of backbone network modules,CA-C3-SPPF,C3-CA-SPPF and C3-SPPF-CA,are designed.For prediction frame fitting,the EIOU loss function is introduced to separately calculate the width-height gap between the prediction frames and the real frames to solve the problem of ambiguity in expressing the width-height ratio in CIOU to the actual width-height gap.The experiments on PASCAL VOC 2007+2012 showed that the m AP of the three improved models improved by 2.2%,2.2% and 3.0%respectively compared to YOLOv5 s.To address the problem that the objects to be measured may appear dense in the image,resulting in poor accuracy of the algorithm in the prediction process,an improved YOLOv5 s algorithm based on feature fusion is proposed.In terms of feature fusion,the improved Bi FPN network is used to replace the path aggregation network to improve the feature fusion capability of different layers in the backbone network.The P2 feature layer,which retains more image information,is added to further improve the algorithm’s ability to process complex information in the image.For the target detection output network,a decoupling detector is used to perform classification tasks and regression tasks through different networks to reduce the mutual influence between different tasks.In addition,by introducing the depth-separable convolution to compress the parameters,the algorithm computation is not significantly increased and the detection speed of the algorithm is guaranteed.Through experiments on the PASCAL VOC 2007+2012 and Wider Person datasets,it is found that the m AP of the improved model is increased by 4.1% and 2.6%,respectively,compared to YOLOv5 s,indicating the effectiveness of the algorithm improvement.