Research On Small Object Detection Algorithm Based On Improved YOLOv5

In recent years,object detection technology has become a hotspot in the research field of computer vision,which is widely used in UAV scene analysis,video surveillance,smart medical care and other fields.Nowadays,the detection of large and medium objects can’t meet the actual needs.People put forward higher requirements for the performance of the algorithm.It is particularly important to improve the detection ability of the algorithm for small objects.However,small objects are difficult to detect because of the following reasons: poor resolution of small objects,dense distribution of tags,large scale difference between objects and susceptibility to background interference.In view of the above challenges,we take YOLOv5 s as the baseline model,and optimize the Head and Neck of YOLOv5 s to get better detection effect.The work can be summarized as follows:(1)In view of the problem of missing detection due to the low resolution of small objects,a P2 detector is added to the Head of YOLOv5 s to detect smaller objects.The model after improvement changes three-scale detection into four-scale detection,which improves the performance of multi-scale object detection.Moreover,the P2 detector combines high-resolution feature maps,which can transfer more shallow features to deep features.In this way,the information of feature maps is enriched,which can effectively improve the problem that small objects at a long distance are easy to miss detection.(2)To solve the problem of interference from image noise and background,the CBAM modules are added to the Neck of YOLOv5 s.The attention mechanism is used to update the feature map before fusion to ignore some irrelevant information and focus more attention on key features.So that the feature map used for object detection after fusion contains more effective information.By adding CBAM module,the antiinterference ability of the network is enhanced,moreover,the important features of the object are strengthened from the channel and space dimensions.It is helpful to enhance the network’s attention to small objects and improve the precision of object detection.(3)In order to enhance the feature processing ability of YOLOv5s’ s Neck and alleviate the problem of insufficient feature information of small objects,we use the idea of Bi FPN to optimize the feature fusion structure of YOLOv5 s.On the basis of the original top-down and bottom-up feature fusion paths,two horizontal cross-scale connections are added.Through the way of efficient bidirectional cross-scale connections,the high-level semantic features in the deep feature map can be fully integrated with the fine-grained features in the shallow feature map,which can enhance the information transmission between different network layers and enrich the feature of small objects.Compared with YOLOv5 s,on the Tiny Person dataset,the YOLOv5s-P2-CBAMBi FPN model has better performance in P,R,m AP@.5 and m AP@.5:.95 increased by0.2%,3.4%,2.6% and 1.13% respectively.On the Vis Drone2019 dataset,the YOLOv5s-P2-CBAM-Bi FPN model increased by 5.1%,5%,5.7% and 3.8%respectively.