Due to the lack of feature information and complex background, the task of small object detection is very challenging. To solve these problems, this paper proposes two small object detection performance enhancement modules for multiple detection tasks and an efficient small object detection network called Multiscale Dilate Feature Enhancement Network (MDFE-Net). MDFE-Net includes two innovative plug-and-play modules: the multi-scale dilated feature aggregation (MDFA) module and the context feature enhancement (CFE) module. MDFA improves the efficiency of multi-scale feature fusion, which is used to capture multi-scale context information and improve the expression of underlying feature information. CFE improves the local feature perception and preserves and extracts the effective information of small image objects to the maximum extent. The network enhances the perception of small objects' feature information and restrains the problem of complex and confusing backgrounds to some extent. We used two public datasets (VisDrone and GTSDB) and a self-built agricultural small object dataset (PSD-Node) to verify the effectiveness of the method. On the above three datasets, the AP50 of MDFE-Net reached 0.304, 0.952, and 0.895, and the AP is 0.172, 0.805, and 0.476, respectively, which exceeded the benchmark model and the current SOTA method.This research presents an innovative small object detection network and provides a reliable technical solution for agricultural small object detection.
Liu et al. (Tue,) studied this question.