The Research On Enhancement Strategy For Small Object Detection

Object detection is an essential branch of computer vision and has been widely used in some regions.At present,most of the practical object detection scenarios are based on regular-sized objects.In contrast,the detection of small objects that exceed the accuracy of human eye recognition could be better,thus making the detection of small objects a difficult task in object detection.With the development of technology,detecting small objects is becoming increasingly widely used,for example,in remote sensing,forest fire prevention,automatic driving,and so on.The relevant features are not noticeable,as small objects occupy small pixels in the video or image.It is challenging to display practical feature information such as appearance,contour,and color,making detecting and effectively capturing useful feature information in small objects difficult.In order to solve these problems,based on the YOLO object detection algorithm,this paper researches the enhancement method for small object detection based on the characteristics of small object detection with few pixels,few feature points,unbalanced samples,and low signal-to-noise ratio,combines the frame synchronization time parameters and the actual application scenarios in the field of remote sensing,and improves the two strategies.These are progressive detection strategy for small objects and remote sensing image object detection strategy.The strategies have been improved.The main research work focuses on the following two aspects.(1)An improved incremental detection strategy for small objects.This strategy is based on the object detection of YOLOv4.It addresses the problems of unbalanced input data samples,small input valid data area,variable input image environment,the low detection accuracy of frame synchronous time parameter detection,etc.The proposed data enhancement method improves the YOLOv4 object detection neural network to achieve progressive detection,effectively enhances the feature information of small objects,and innovatively proposes the conversion algorithm from position coordinates to frame synchronization time parameters.The experimental results show that the proposed strategy has higher accuracy than other popular strategies and can automatically calculate the frame synchronization time parameters.(2)A remote sensing image object detection strategy for small objects.This strategy is based on the object detection of YOLOv7-tiny.It addresses the problems of small object feature information in the input image,large full-image size,and significant noise in remote sensing imaging by using an improved attention module and proposing an enhanced neural network YOLOv7a-tiny and loss function to enhance the learning ability of small objects.The experimental results show that the proposed strategy can achieve higher detection accuracy and faster convergence speed than YOLOv7.In summary,this paper is oriented to the study of small object detection enhancement strategies based on the YOLO series of object detection strategies for a series of problems faced by small object detection,improve the corresponding object detection neural network and attention module,improve the small object detection accuracy,for small object detection in remote sensing,automated driving and other fields of application has a vital reference role.