Infrared small target segmentation with multiscale feature representation

• Published in: Infrared physics & technology Vol. 116; p. 103755
• Main Authors: Huang, Lian, Dai, Shaosheng, Huang, Tao, Huang, Xiangkang, Wang, Haining
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2021
• Subjects: Attention; Convolutional neural networks; Infrared image; Local similarity; Small target segmentation; Spatial pyramid; Local similarity; Spatial pyramid; Convolutional neural networks; Attention; Infrared image; Small target segmentation
• ISSN: 1350-4495; 1879-0275
• DOI: 10.1016/j.infrared.2021.103755

•The local similarity pyramid module effectively captures multiscale features of infrared small targets.•The feature aggregation module considers to merge shallow and deep features with attention.•The proposed network outperforms other state-of-the-art methods.•The ablation study demonstrates the contribution of each module of proposed network. Small target segmentation is one of the vital techniques in various infrared-based applications. The typical challenges are summarized as follows: the sizes of infrared small target are extremely small compared with common targets, and infrared small targets with dim appearances are similar to the background noise. To address the above problem, this paper studies how to leverage the powerful pyramid structure and attention mechanism for the segmentation of infrared small targets. Multiple well-designed local similarity pyramid modules (LSPMs) are endowed with a strong capability to model the multiscale features of infrared small targets. Specifically, each LSPM with a different scale estimates the weight of the local similarity, which quantifies the degree to which a pixel is similar to other pixels. The pyramid features are introduced into the feature aggregation module as the supplement of the global features. The proposed network aggregates features with different weights that facilitate the fusion of shallow and deep features. We empirically evaluate the proposed network on public infrared small target segmentation datasets. The experimental results demonstrate that the network achieves better performance than other state-of-the-art methods. The code is publicly available at https://github.com/HuangLian126/LSPM.