Detection of Photoelectric Small Targets in Unstable Videos Inspired by the Visual Dynamic Information Processing Mechanism

Small target detection and recognition in complex backgrounds based on photoelectric video is a key technology in the field of security. However， due to factors such as carrier motion and external disturbances， photoelectric detection imaging often suffers from unstable jitter. This jitter disrupts the temporal consistency of the background， reduces the signal-to-noise ratio （SNR） of small targets， and degrades detection performance. To address the degradation of photoelectric small target detection performance under unstable video conditions， this study proposes a bio-inspired photoelectric small target detection architecture based on the dynamic information processing mechanism of biological vision. First， an image stabilization preprocessing method for unstable imaging sequences is designed. By suppressing global background interference and enhancing the temporal response characteristics of small targets， this method provides more stable input signals for subsequent feature extraction. Second， inspired by the dynamic information processing mechanism of the magnocellular pathway in the biological visual system， a motion response feature map is constructed. This map is combined with stabilized video frames to enhance the motion saliency and regional discriminability of small targets. Furthermore， the temporal motion information extracted after image stabilization and image appearance features are jointly modeled within a convolutional neural network. This modeling serves as an important supplementary signal to guide the deep detection network， thereby constructing the bio-inspired detection model named SMSOD. This model effectively enhances the perception capability of infrared small targets in complex dynamic backgrounds. Experimental verification is conducted on real photoelectric small target datasets. The detection accuracy under unstable conditions reaches 73.1%. After image stabilization preprocessing， the detection accuracy increases to 74.9%. With further fusion of motion information， the detection accuracy reaches 91.8%， representing an improvement of 18.7% compared with the detection results obtained from the original jittery frames. The results demonstrate that the proposed method can effectively suppress the adverse effects caused by global jitter， providing a stable and robust signal processing approach for small target detection under unstable video conditions.