Toward Accurate Camouflaged Object Detection With In-Layer Information Enhancement and Cross-Layer Information Aggregation

According to the process of the human visual observation mechanism we propose a novel architecture for camouflaged object detection (COD), which is composed of two main components: 1) in-layer information enhancement module (IIE) and 2) cross-layer information aggregation module (CIA). Specifically, the IIE module is to simulate the first step of human eyes detecting the target, namely, judging whether the camouflaged object exists and roughly be located. Besides, the CIA module is leveraged to imitate the second process of the human visual observation mechanism, refining the edge of the camouflaged object and eliminating interference. Since the shallow texture information and the deep semantic information are complementary, we combine the in-layer information with the cross-layer information to more accurately locate the target object and avoid noise and interference at the same time. Our model outperforms 13 state-of-the-art deep learning-based methods upon three public data sets in terms of four widely used metrics.