KDSMALL: A lightweight small object detection algorithm based on knowledge distillation

With the development of computer vision, small object detection has become a research pain point and difficulty in computer vision. Feature acquisition and accurate localization of small objects are two serious challenges that exist for small objects at present. In this paper, a generalized small object detection algorithm is formed based on a multi-scale feature extractor, a feature search network with hybrid attention mechanism, and knowledge distillation. The algorithm firstly performs feature extraction of small objects based on multi-scale feature extractor, secondly uses CBAM attention mechanism and Efficient network to perform feature search on features obtained from the feature map to help obtain more features of the small object, and finally performs knowledge distillation on the baseline model based on the idea of teacher–student knowledge distillation to help the baseline model locate the detected object. In this paper, YOLOv5s is selected as the benchmark experiment, and the designed algorithm is fused to YOLOv5s, compared with the baseline model, the fused model's experimental metrics mAP on the VOC mixed dataset is improved by 14.45% on average. The experimental results show that the designed algorithm can effectively improve the detection performance of the object detection model for small objects.