Research And Application Of Small Object Detection Algorithm Based On YOLOv5

With the popular application of artificial intelligence technology such as automatic driving and UAV identification,computer vision has become one of the most popular research fields at present.As the core problem in the field of computer vision,the application demand of object detection is increasing.Target detection has always been a challenging problem in the field of computer vision,especially for small target detection scenes.This is because the number of small targets in the original image is small and the pixels occupied are small,which leads to fewer features that the detector can extract,thus affecting the detection effect of small targets.The main research work of this thesis is as follows:(1)Improve the network structure to enhance the detection effect of the model on small target objects from the input side,Backbone,Neck,Prediction and other aspects.(2)Add small target detection layer,and get appropriate modification parameters after several experiments.(3)Add segmentation calculation framework to solve the problem of insufficient detection accuracy of small and medium-sized objects in high-resolution images.(4)The YOLOv5-Sobj model was finally obtained by analyzing the experimental results.In this thesis,an improved small-target detection algorithm based on the one-stage detection algorithm YOLOv5 is proposed to improve the detection performance of small objects.To achieve this,this article will examine how replacing or modifying certain structures and other factors in the original model can affect detection performance and reasoning time.Therefore,this thesis proposed a series of models with different scales and named them "YOLOv5-Sobj",and set up relevant experiments to detect the improvement effect.Experimental results show that when 50% Io U is used to detect smaller target objects,the m AP of the model can be improved by about 5.3%,but the cost is that compared with the original YOLOv5 model,his reasoning time is increased by 4ms.The purpose of this study is to provide insight into how certain changes affect small target detection,and the findings can be applied in future environments where small target detection is needed,such as autonomous driving.