Two-Stage Polishing Network for Camouflaged Object Detection

Abstract

Due to the needs of medical and military fields, Camouflaged Object Detection (COD) becomes one of important branches of object detection. It has gradually gained people’s attention in recent years. How to correctly locate camouflaged objects accurately segment them are the main problems in this field. The COD task is far different from SOD because of the more complex background with similar colors and textures. Therefore it is more challenging. At present, there are still few existing methods, and lacks targeted method for the edge detection problem. In this paper, we propose a novel Two-Stage Polishing Network (TSPNet). This network consists of Front Feature Fusion Module (FFFM) and Polishing Module (PM). FFFM adopts Cross-modal Feature Aggregation and Global and Local Feature Aggregation to capture global context information and detail local information simultaneously. Meanwhile, PM uses edge truth map as supervision imformation to further study the object edge features. The experiments was conducted on three available datasets and the result shows that the proposed framework outperforms state-of-the-arts. Besides, TSPNet is compact with 50% model size saving than existed COD models.