Camouflaged Object Segmentation Based on Joint Salient Object for Contrastive Learning

Abstract

In a broad sense, camouflaged objects generally refer to objects that have a high degree of similarity to the background. Therefore, camouflaged object segmentation (COS) is more challenging than traditional object segmentation. Current COS networks have high segmentation precision on datasets. However, the problems of object miss detection and false alarm still occur, mainly due to the different camouflage levels of the objects. In this paper, we propose a Joint Comparative Learning Network (JCNet) for camouflaged object segmentation based on joint salient object for contrastive learning to overcome the widespread challenges in COS. Specifically, the main innovation of JCNet is the Contrastive Network (CNet) design, which generates a unique feature representation of the camouflaged object different from others. In terms of details, we design an edge guidance module to enhance the edge extraction capability. Moreover, a global relationship capture module is proposed to improve the confidence level of the feature representation. Finally, we set positive and negative samples and loss functions in conjunction with sample types. We conducted comprehensive experiments using four COS datasets, and the results demonstrate its suitability for COS when compared with other state-of-the-art segmentation models. JCNet achieves optimal results on five evaluation metrics, including an average improvement of 1.93% and 2.7% on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F_m</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</i> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ω</sup> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><i>m</i></sub> , respectively. In summary, it has lower miss and false alarm rates, and better generalization in the COS task. In addition, the experiments demonstrate that JCNet also has strong segmentation capability in salient object segmentation, achieving a win-win situation for both tasks. The code will be available at https://github.com/jiangxinhao2020/JCNet.