UNETCOD: AN EFFICIENT UNET3+ FRAMEWORK WITH DUAL-ATTENTION AND BOUNDARY ENHANCEMENT FOR CAMOUFLAGE OBJECT SEGMENTATION

Abstract

Camouflaged object segmentation (COS) is challenging due to blurred boundaries, low contrast, and complex backgrounds, while practical deployment often requires lightweight models suitable for edge devices. In this work, we conduct a systematic study of UNetbased architectures, including UNet, UNet++, and UNet3+, for COS using lightweight encoder backbones, with a focus on camouflaged targets in complex real-world environments. Starting from ResNet-based encoders, we further investigate EfficientNet variants from B0 to B5 to achieve a better balance between accuracy and computational efficiency.

To enhance boundary discrimination in camouflage scenarios, we integrate a Convolutional Block Attention Module (CBAM) after each encoder stage to refine multi-scale features, combine it with a Boundary Enhancement Module (BEM), and propose a warm-up-based training strategy that introduces a boundary loss after an initial stabilization phase using only segmentation loss. Experimental results on the MHCD-Seg dataset show that UNet3+ with an EfficientNet-B3 backbone and BEM, combined with a 20-epoch warmup strategy, delivers the best performance among all evaluated configurations, providing an effective solution for camouflage object segmentation on resource-constrained edge devices. The implementation code has been released and is publicly accessible at https://github.com/hangtruong85/UNetCOD.