Location-Guided Dense Nested Attention Network for Infrared Small Target Detection
Huinan Guo,Nengshuang Zhang,Jing Zhang,Wuxia Zhang,Congying Sun
DOI: https://doi.org/10.1109/jstars.2024.3472041
IF: 4.715
2024-10-25
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Abstract:Infrared small target (IST) detection involves identifying objects that occupy fewer than 81 pixels in a 256 × 256 image. Because the target is small and lacks texture, structure, and shape information on its surface, this task is highly challenging. CNN-based methods can extract rich features of the target. However, overly deep network structures may increase the risk of losing small targets. In addition, pixel-level positional deviations can also reduce the detection accuracy of IST. To address these challenges, we propose the location-guided dense nested attention network for IST detection. The proposed network consists of a pixel attention guided feature extraction module (PAG-FEM), a channel attention guided feature fusion module (CAG-FFM), and a detection module. First, the PAG-FEM utilizes the DNIM dense nested blocks from the DNANet as the backbone, integrating both channel and pixel attention mechanisms. This method focuses on the semantic and positional information of the targets, yielding semantic features that emphasize the positions of small targets. Second, the CAG-FFM employs upsampling and convolution operations to align the feature sizes, while utilizing the channel attention mechanism to obtain effective channel information. Then, these features are fused through stacking, addition, and averaging operations to obtain more discriminative features. Finally, the detection module uses eight-connected neighborhood clustering method to obtain the centroid coordinates of the targets for subsequent detection evaluation. Three datasets are utilized to verify our method, and experimental results show that our method performs better than other advanced methods.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geography, physical