High-Precision and Fast Inference for Infrared Small Target Detection through Semantic Gap Reduction

In recent years, significant advancements have been made in the field of infrared small target detection (IRSTD), largely driven by developments in deep learning and computer vision. Deep learning-based methods have demonstrated substantial improvements in both accuracy and inference speed compared to traditional approaches, enabling their integration into real-time embedded systems. However, many data-driven techniques rely on complex network architectures to process large volumes of intricate data, resulting in additional computational overhead. To enhance the efficiency of IRSTD, we propose an improvement based on the classical segmentation framework, introducing a semantic gap elimination module (SGEM) to reduce the level-to-level semantic gap. This enhancement improves the stability and performance of IRSTD. Notably, our method does not rely on complex network architectures, allowing it to outperform other deep learning-based methods in terms of computational efficiency. It also exceeds the performance of the fastest methods, achieving more than a threefold increase in the frames per second (FPS). Furthermore, comparative experiments demonstrate the effectiveness of our approach, showing superior performance over recent methods in both segmentation and localization accuracy.