Improving small object detection via context-aware and feature-enhanced plug-and-play modules

Detecting small objects is a challenging task in computer vision due to the objects only occupying a limited number of pixels and having blurred contours. These factors result in minimal discriminative features being available to effectively model the objects. In this paper, we propose three lightweight plug-and-play modules that can be seamlessly integrated into object detection algorithms, particularly those in the YOLO series, to improve the accuracy of detecting small objects. The Spatially Enhanced Convolutional Block Attention Module (SE-CBAM) is integrated into the feature extraction layer of the network to enhance the feature extraction capability of neural networks. Additionally, a Contextual Information Pooling Enhancement Module (CIE-Pool) is included at the multi-scale feature fusion stage to extract and improve object background information, which enhances the recognition rate of small objects. To improve the detection of small objects, a new layer is added to the detection head, which incorporates the shallow feature map obtained from the feature extraction network after Adaptive Feature Processing (AFP), thereby obtaining more and richer information about small objects. The efficacy of the algorithm has been evaluated on the VisDrone2021 and AI-TOD datasets. The experimental results demonstrate that the method proposed in this paper greatly improves the detection accuracy of small objects while maintaining real-time capabilities. Furthermore, it maintains high accuracy and speed even when dealing with complex background conditions and detecting small objects with high blur.