An Information Calibration and Sliding Mining Network for Co-Saliency Object Detection
Longsheng Wei, Jiu Huang · IEEE Transactions on Instrumentation and Measurement · 2024
Many co-saliency object detection methods do not consider the information loss caused by different sizes or numbers of channels in feature fusion. In addition, some feature extraction modules bring improved accuracy while ignoring local information and it is computationally heavy. In this article, we propose a novel end-to-end information calibration and sliding mining co-saliency detector (ICSM) for co-saliency object detection. In the stage of encoding features and decoding features fusion, we propose a multistage calibration correlation (MCC) module. In the MCC module, we calibrate the information from the features of different sizes and channel numbers to reduce the information loss when these features are fused, where a series of region-level modeling is carried out to preliminarily calculate the correlation between image pixels. Then, we propose a sliding region mining (SRM) module, which converts pixel-level features into foreground and background contrast token features. It uses these token features to search for synergistic saliency objects across multiple images and to suppress background interference while reducing computing costs. Extensive experiments show that our model achieves better performance than the excellent co-saliency object detection methods in recent years under the three most popular benchmark datasets.