Abstract:

Infrared Small Target Detection is a critical focus of various fields,including earth observation and disaster relief efforts,receiving considerable attention within the academic community.Since infrared small targets often occupy just a few dozen pixels and are scattered within complex backgrounds,it becomes paramount to extract semantic information from a broad range of image features to distinguish targets from their surroundings and enhance detection performance.Traditional convolutional neural networks,due to their limited receptive fields and substantial computational demands,face challenges in effectively capturing the shape and precise positioning of small targets,leading to missed detections and false alarms.In response to these challenges,this paper proposes a novel Smooth Interactive Compression Network comprising two main components:the Smooth Interaction Module and the Cross Compression Module.The Smooth Interaction Module extends the feature map's receptive field and enhances inter-feature dependencies,thus bolstering the network’s detection robustness in complex background scenarios.The Cross Compression Module takes into account channel contributions and the interpretability of pruning,dynamically fusing feature maps of varying resolutions.Extensive experiments conducted on the publicly available SIRST dataset and IRSTD-1K dataset demonstrate that the proposed network effectively addresses issues such as target loss,a high false alarm rate,and subpar visual results.Taking the SIRST dataset as an example,compared to the second-best performing model,the proposed model achieved a remarkable improvement in metrics:IoU,nIoU,and P_d are increased by 3.05%,3.41%,and 1.02%,respectively.Meanwhile,F_a and FLOPs are decreased by 33.33% and 82.30%,respectively.

Key words: Infrared Small Target Detection, deep learning, network coding, model compression