Enhancing Small Object Detection via Differential Convolution Fusion Modules

Small object detection, as a branch of object detection tasks, often exhibits unsatisfactory performance due to challenges such as detecting small objects against cluttered backgrounds. Inspired by the reliance of the human visual system on the contrast between objects and their surrounding environments in small object detection, this paper proposes a method utilizing differential convolution to highlight the differences between small objects and their surroundings, and constructs a corresponding novel feature enhancement module. Specifically, based on the analysis of several existing differential convolution methods, we first discuss their advantages and disadvantages, and propose the Differential Convolution Module (DCM) to emphasize the center-surround contrast of small objects. DCM combines three types of differential convolutions to effectively enhance feature maps. Furthermore, by fusing original features with such enhanced features, we are finally able to enhance the representation of target features and improve the accuracy of object detection. In addition, DCM is featured with simplicity in computation, good interpretability, and easy integration into the existing popular object detection models. The effectiveness of the proposed method is validated on the VisDrone2019 [1] object detection dataset, and experimental results demonstrate significant performance improvements in small object detection tasks.