Abstract
Small objects occupy a very small proportion of pixels in images, making feature extraction difficult. This leads to existing neural networks struggling to accurately determine the location and category of small objects. Meanwhile, it is difficult to manually label the position and category of small objects, which makes the small objects lack of enough high-quality datasets. Most of the existing solutions rely on additional auxiliary branches to focus on the enhancement of spatial domain features, the model generalization ability is weak and the detection performance is heavily dependent on high-quality data sets. Therefore, we design a semi supervised detection algorithm KPN (Keypoint Network) for small targets, which uses custom keypoints to solve the problem of small target localization with limited labeled data. Specifically, we rethink the definition and role of keypoints, and thus define three types of keypoints to guide the network in accurately localizing small objects: the object center point for bounding box positioning, the object presence point with the highest probability in the heatmap and the adaptive point with strong response from the backbone network. Additionally, to address the difficulty of feature extraction of small targets, we introduce the frequency domain supplement module to focus on the frequency domain feature enhancement. Finally, we propose a joint confidence estimation and an adaptive thresholding strategy to resolve pseudo-label ambiguity. The experimental results confirm the superiority of our approach against other advanced methods, showing significant performance gains in small object detection.