SPSDet: Advances in Infrared Small-Scale Ship Detection Using State Space Models

Besides small vessel target detection in maritime images, other computer-vision-related challenges also remain-primarily associated with limited data concerning small targets as they are often obscured by larger objects or background noise. Therefore, Forest-Transformer-based models tend to struggle with the size of specific datasets impacting on an effective response to targets of various orientation and scales. Hence, some models need to work with more agility and less cumbersome to address these issues. In this respect, this paper proposes exciting two methods improving ocean craft detection and segmentation significantly. The first is that the lightweight YOLO v9 architecture is integrated into the SliceBoost framework using programmable gradient information to minimize loss of critical information during feature extraction. The second is the Vision Mamba model that further augments spatial awareness by embedding position information into a newly constructed bidirectional state-space model that enhances contextual modeling abilities. To adept at the modeling of small vessel detection in infrared maritime images, it combines the local feature processing efficiency of convolution-only architectures with the global knowledge aberration capabilities of Transformer-based architectures. Experimental results show significant improvements in both detection and processing efficiency making real-time small-target detection possible within varying maritime applications. Furthermore, the paper discusses possible future maritime target recognition applications of these two techniques. The source code is available at the following link: https://github.com/CaiGuoHui123/SPSDet.