Small object detection by Edge-aware Neural Network

The object detection method is widely applied in industrial inspections. However, many detectors face challenges in accurately capturing the blurred edge details of small objects, resulting in inaccurate bounding box predictions. To address this, we propose an Edge-aware Neural Network (EANN) for small object detection. Firstly, we introduce a Channel and Spatial Attention Fusion Module (CSAFM) to enhance the edge features of small objects, enabling the network to extract more discriminative information. Next, we propose a Multiple Aggregation Feature Pyramid (MAFP) to integrate multi-scale deep features into shallow features. This fusion enriches the shallow features with abundant semantic information, thereby aiding in the detection of small objects. Additionally, we propose a Side and Center Point Aligned Intersection over Union loss (SCPAIoULoss) to enhance the bounding box regression when there is minimal overlap between predicted and ground truth boxes. SCPAIoULoss combines Side Ratio (SR) loss, Center Point Distance (CPD) loss, and Intersection over Union (IoU) loss. The utilization of SR Loss directly constrains the width and height regression of bounding boxes, while CPD loss introduces stricter constraints to facilitate bounding box regression. Furthermore, IoU loss promotes the overall regression of predicted boxes. We extensively experiment on Tiny CityPersons, WiderFace, and our proposed dataset of base station data centers to validate the effectiveness of our method. The results indicate that our method surpasses several State-of-The-Art (SOTA) approaches in small object detection and can be effectively applied to the task of inspecting base station data centers.