文献类型：期刊文献

英文题名：CourtNet: Dynamically balance the precision and recall rates in infrared small target detection

作者：Peng, Jingchao[1];Zhao, Haitao[1];Zhao, Kaijie[1];Wang, Zhongze[1];Yao, Lujian[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Automat Dept, Meilong Rd 130, Shanghai 200237, Peoples R China

年份：2023

卷号：233

外文期刊名：EXPERT SYSTEMS WITH APPLICATIONS

收录：;EI(收录号：20233014442546);WOS:【SCI-EXPANDED(收录号:WOS:001048866000001)】；

基金：Acknowledgments This work was supported by the National Natural Science Founda-tion of China (NSFC) under Grant 62173143 and Grant 61973122.

语种：英文

外文关键词：Target detection; Infrared small target detection; Attention mechanism; Transformer

摘要：Infrared small-target detection (ISTD) is an important computer vision task. ISTD aims at separating small targets from complex background clutter. The infrared radiation decays with distance, making the targets highly dim and prone to confusion with the background clutter, which makes the detector challenging to balance the precision and recall rates. To deal with this difficulty, this paper proposes a neural-network-based ISTD method called CourtNet, which has three sub-networks: the prosecution network is designed to improve the recall rate; the defendant network is devoted to increasing the precision rate; the jury network weights their results to adaptively balance the precision and recall rates. CourtNet takes the structure of Transformers, whose feature resolution remains unchanged. Furthermore, the prosecution network utilizes a densely connected structure, which can prevent small targets from disappearing in the forward propagation. In addition, a fine-grained attention module performs attention inside patches to accurately locate the small targets. This paper implements extensive experiments on two ISTD datasets, MFIRST and SIRST, and compares CourtNet with ten other traditional and deep-learning-based methods. Experimental results show that with the fast detection speed (60.61 FPS), CourtNet achieves the best F1 score, 0.62 (in MFIRST) and 0.73 (in SIRST), among the compared methods. The code and dataset will be available at https://github.com/PengJingchao/CourtNet.