FATCNet : Feature Adaptive Transformer and CNN for Infrared Small Target Detection

Infrared small target detection (IRSTD) is the process of separating targets from complex and blurry backgrounds, with wide applications in rescue, maritime guidance, and other fields. However, traditional deep learning methods that rely solely on CNN-based methods are prone to losing targets as the convolution deepens, resulting in missed detections when segmenting small infrared targets. Similarly, because of the unique characteristics of infrared small targets, Transformer-based methods lack the supplement of convolutional local detail features, making it difficult to achieve good results. To address these issues, we propose the Feature Adaptive Transformer and CNN network (FATCNet): a dual branch decoder structure network that utilizes CNN and Transformer to obtain local and global features in parallel. Meanwhile, we propose the Attention Guided Adaptive Fusion Block (AGAF) to better integrate global and local feature information, eliminate semantic gaps, and dynamically fuse features based on network learning. The network encoder incorporates the Hierarchical Attention Module (HAM), which leverages the distinct characteristics of various hierarchical feature maps to enhance features, thereby mitigating the risk of losing crucial details pertaining to infrared small targets. Additionally, We have constructed a bird infrared dataset, named BSIRST v2, which comprises 1000 authentic images with diverse target shapes, sizes, and cluttered backgrounds, and the dataset includes meticulous pixel-level annotations. Our approach's efficacy has been validated through experiments conducted on both public datasets and BSIRST v2. In comparison to other state-of-the-art methods, our FATCNet outperforms in terms of probability of detection (PD), F1 score (F1), and intersection over union (IoU).