Fast Small Object Detection Algorithm Based on Feature Enhancement and Reconstruction

Small object detection is one of the research hotspots and challenges in the field of machine vision. Because small targets have fewer pixels in the image, the number of features captured by the first few layers of the target detection network is very small, and the high-level semantic information such as feature classification disappears quickly, making the accuracy of target detection relatively low. The paper uses sub-pixel convolution to reconstruct feature information, uses SE (Squeeze-and-Excitation) attention mechanism to enhance features, and uses U-Net network to capture feature context information and has the advantages of faster detection speed. The algorithm realizes the fast and accurate identification of small targets. Numerical experimental results on the COCO data set and the SUN data set show that the algorithm improved performance by more than 12% in the case of the bounding box AP compared with mainstream algorithms such as YOLOv4.