Research on training-free small object detection methods driven by visual priors

Abstract

Small objects in long-range surveillance, UAV aerial imagery, infrared weak-target detection, and online industrial inspection generally have a few pixels. They have poor borders and are expensive to acquire as annotated samples. Based on the above reasons, a training-free small-object detection method using visual priors is put forward here. No target category samples are used for network training in this way. A candidate response map is built by combining scale-space local contrast, gradient edges, colour or grayscale rarity, morphological continuity and image-quality constraints. Adaptive thresholding, connected-component filtering and duplicate-box suppression are then employed to find small objects. It is to be applied to scenes with few or no labels, and interpretable detection results are required promptly. Experiments on visible-light, low-contrast and infrared weak-target images show that the proposed method improves recall and reduces false positives compared with traditional saliency and morphological baselines without task-specific fine-tuning. ASimple and easy-to-use solution for rapid engineering deployment and parameter tuning.