WGRNet: Wavelet and Gradient Guided Refinement Network for Multi-Class Military Camouflage Object Segmentation

Camouflaged object detection is a critical problem in military applications where soldiers, tanks, and trucks are delib
erately designed to blend into their surroundings. Traditional computer-vision methods fail when the contrast between 
the object and background is extremely low. This work addresses the problem of multi-class semantic segmentation 
of camouflaged military targets across three foreground classes (Soldier, Tank, Truck) using deep learning. Five deep
learning segmentation architectures were implemented and compared: SCPNet, SINetV2, DGNet, EW-CASCADE, 
and our proposed WGRNet. All models were trained on a three-class subset of the self-constructed MilCamSeg dataset 
using a common loss function (Focal Loss + Dice Loss) to ensure a fair comparison. WGRNet is a custom architecture 
that combines the PVTv2-B2 transformer backbone, wavelet-based frequency decomposition, and gradient-weighted 
attention, specifically designed to handle the low-contrast patterns found in camouflage. Experimental results show 
that EW-CASCADE achieves the best mean IoU (0.821) and pixel accuracy (0.962), while WGRNet achieves the best 
Dice overlap (0.896). Both models significantly outperform the other three architectures on all foreground classes, 
demonstrating that transformer backbones combined with wavelet- and attention-based decoding are highly effective 
for camouflaged object segmentation.