Mscnet: Mask stepwise calibration network for camouflaged object detection

Abstract

Camouflaged object detection (COD) aims to accurately segment camouflaged objects blending into the environment and is a challenging task. Most existing deep learning-based COD methods do not explicitly enhance the region information of camouflaged objects, nor do they use the region information for mask calibration. To solve this issue, we propose a novel mask stepwise calibration network (MSCNet) for camouflaged object detection, which achieves high-precision detection of camouflaged objects. Specifically, MSCNet consists of a region information enhancement encoder and a mask stepwise calibration decoder. In the encoder, we first utilize a PVT backbone network to extract different levels of features from camouflaged images. Then, we design a region information enhancement module to enhance the region information of camouflaged objects while suppressing the interference of background information by mining, embedding, and aggregating the region information in different levels of features. In the decoder, we first design a coarse mask generation module to generate coarse prediction masks of camouflaged objects by directly cross-fusing different levels of features extracted by the backbone. In addition, we also design a mask calibration module to calibrate coarse prediction masks of camouflaged objects using the region information of different levels of camouflaged objects as a guide. Extensive experimental results on four benchmark datasets show that our method effectively identifies camouflaged objects and surpasses most state-of-the-art COD methods.