ESOD-YOLOv8: Small object detection enhanced with auto-disturbance rejection convolution
Zhenhua Yu, Huize Liang, Ou Ye, Yun Zhang
Abstract
Small objects are significantly disturbed by the background and prone to losing features during the sampling process. To address these problems, we propose an optimization method of YOLOv8 based on Auto-Disturbance Rejection Convolution (ADRConv). Firstly, we design a new ADRConv, which uses the quadratic central difference to extract small target features and combines ordinary convolution to represent the effective background. Subsequently, we construct a feature extraction selector within the YOLOv8 network, and the activation strategy function controls the difference operation to ensure a balanced representation of object features and background. ADRConv is also applied to enhance the feature fusion network of the YOLOv8 structure, addressing the challenge of feature loss associated with small objects. The experimental results show that compared with YOLOv8 benchmark algorithm and mainstream methods in recent years, the precision of this proposed method is increased by 4.3 % and 2.6 % , the recall rate is increased by 7.5 % and 0.9 % , and the average detection accuracy is improved by 3.9 % and 1.3 % , which effectively enhances the detection effect of small targets.