BiCOD: A Camouflaged Object Detection Method Directed by Cognitive Attention

• Published in: IEEE sensors journal 2024-02, Vol.24 (4), p.4711-4721
• Main Authors: Xu, Lianrui, You, Xiong, Jia, Fenli, Liu, Kangyu
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Attention; camouflaged object detection (COD); Closed loops; Cognition; Computational modeling; Eye movements; Feature extraction; Frequency-domain analysis; Image segmentation; Machine learning; Modules; Object detection; Object recognition; Task analysis; Visualization
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2023.3343917

Camouflaged object detection (COD) is a typical application of deep-coupled unmanned platform combat support, which aims to detect objects that are highly similar to the background in terms of structure, details, and texture while improving the efficiency and accuracy of detecting camouflaged objects. The existing COD methods are built upon extraction and segmentation of image features and lack of theoretical interpretability. In this article, the task of COD was revisited and analyzed. From the perspective of cognition, the cognitive laws of camouflaged objects were assessed through eye movement experiments to form an entire cognitive process, which serves as a guide for designing COD methods. Feature extraction, position attention, and channel attention modules (CAMs) were utilized as the basic framework. The residual-in-residual module was introduced to improve the accuracy of feature learning and transmission. Then, a bidirectional attention module (BAM) was added to guide the feedforward and feedback of attention features, and a closed loop was formed to achieve efficient feature transmission and use. As a result, the performance of our method BiCOD was promoted. In addition, a COD dataset containing both natural and artificial camouflage objects was compiled to evaluate the generalization ability of the camouflaged object recognition algorithm. The experimental results showed that BiCOD achieved an advanced level in quantitative results and visual comparisons in general, and the effectiveness and accuracy of the method in different environments were verified.