CubeNet: X-shape connection for camouflaged object detection

• Published in: Pattern recognition Vol. 127; p. 108644
• Main Authors: Zhuge, Mingchen, Lu, Xiankai, Guo, Yiyou, Cai, Zhihua, Chen, Shuhan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2022
• Subjects: Camouflaged object detection; Edge guidance; Neural network; Novel feature aggregation; Novel feature aggregation; Neural network; Edge guidance; Camouflaged object detection
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2022.108644

•We propose a novel CubeNet architecture for camouflaged object detection, which accompanies with feature Fusion Blocks and X-connection to sufficiently integrate multiple layer features.•The proposed model can be trained quickly. Meanwhile, it achieves real-time inference efficiency.•Extensive results on three challenging datasets verify the effectiveness of the proposed method. Camouflaged object detection (COD) aims to detect out-of-attention regions in an image. Current binary segmentation solutions fail to tackle COD easily, since COD is more challenging due to object often accompany with weak boundaries, low contrast, or similar patterns to the background. That is, we need a more efficient scheme to address this problem. In this work, we propose a new COD framework called CubeNet by introducing X connection to the standard encoder-decoder architecture. Specifically, CubeNet consists of two square fusion decoder (SFD) and a sub edge decoder (SED). The special designed SFD takes full advantage of low-level and high-level features extracted from encoder-decoder blocks, providing more powerful representations at each stage. To explicitly modeling the weak boundaries of the objects, we introduced a SED between the two SFD. With such kind of holistic designs, these three decoder modules resolve the challenging ambiguity of camouflaged object detection. CubeNet significantly advance the cutting-edge model on three challenging COD datasets (i.e., COD10K, CAMO, and CHAMELEON), and achieves the real-time (50fps) inference.