# Deep Texture-Aware Features for Camouflaged Object Detection

**Authors:** Jingjing Ren, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Yangyang Xu, Weiming Wang, Zijun Deng, Pheng-Ann Heng  
**arXiv:** [2102.02996](https://arxiv.org/abs/2102.02996)  
**Date:** 2021-02-05  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2102.02996v1/

## Key points

- The paper addresses the challenge of detecting objects with similar textures to their surroundings
- The authors propose a deep convolutional neural network architecture with texture-aware refinement modules (TARMs)
- TARMs compute covariance matrices of feature responses to extract texture information
- An affinity loss is used to separate texture between objects and background
- A boundary-consistency loss helps explore object detail structures
- The proposed approach outperforms state-of-the-art methods in camouflaged object detection
- Incorporating residual refinement blocks (RRBs) improves detection performance
- Removing the boundary-consistency loss leads to inferior results, highlighting its importance
- The paper presents a deep network architecture that learns deep texture-aware features for camouflaged object detection.

## Summary

The paper titled "Deep Texture-Aware Features for Camouflaged Object Detection" addresses the challenging task of detecting objects that have similar textures to their surroundings. The authors propose a novel deep convolutional neural network architecture that aims to amplify the subtle texture differences between camouflaged objects and the background. To achieve this, the authors introduce multiple texture-aware refinement modules (TARMs) within their network. These TARMs compute covariance matrices of feature responses to extract texture information and design an affinity loss that helps separate the texture between camouflaged objects and the background. Additionally, a boundary-consistency loss is adopted to explore object detail structures. The effectiveness of the proposed approach is evaluated on benchmark datasets for camouflaged object detection both qualitatively and quantitatively. The experimental results demonstrate that their method outperforms various state-of-the-art methods by a large margin. Further analysis is conducted to evaluate specific components of the proposed approach. The results show that incorporating residual refinement blocks (RRBs) improves the detection performance compared to a basic network architecture. Moreover, removing the boundary-consistency loss from TARMs leads to inferior results, highlighting its importance in enhancing consistency across boundaries. In conclusion, this paper presents a deep network architecture that learns deep texture-aware features for camouflaged object detection. The proposed approach effectively amplifies texture differences between objects and backgrounds, leading to superior detection performance compared to existing methods. The incorporation of RRBs and boundary-consistency loss further enhances the accuracy of detecting camouflaged objects in complex situations.