Infrared Small Target Detection with Wavelet Attentions

Recently, the field of infrared small target detection (IRSTD) has seen a significant shift towards deep learning methodologies. While these techniques have achieved notable success, they predominantly emphasize the development of complex model architectures aimed at extracting more distinctive spatial features, frequently neglecting the importance of frequency-domain characteristics. This imbalance in feature extraction versus the overall effectiveness of small target detectors, as current deep learning methods excel in capturing spatial details but are less adept at recognizing frequency-based attributes. To tackle this issue, our work introduces an innovative approach that boosts detection accuracy by incorporating high-frequency components, which are extracted using 2-D discrete wavelet transforms. A key component of our solution is the introduction of a novel block, High-Frequency Attention (HFA). The HFA block serves two primary functions: it acts as a substitute for conventional pooling layers and reduces information loss typically associated with standard (maximum or average) pooling operations by integrating high-frequency elements post-pooling. By implementing the HFA block in place of old-styled pooling layers at the input stage of each feature extractor, our proposed HFA-Net demonstrates superior performance compared to leading contemporary methods, achieving marked enhancements in detection precision.