Deep SINET-based Framework for Camouflaged Object Detection in Military Environments with Backbone Performance Analysis

Abstract

In this paper, propose an updated and experimentally consistent Deep SINet-based architecture for the military camouflaged item detection task with the ACD1K dataset, which includes 1,000 annotated military camouflage photos with 700 training images and 300 testing images. In contrast to the wide application of SINet, our work focuses on a controlled backbone-performance examination, where VGG19, ResNet-50, and ResNet-101 are incorporated into the same SINet search-identification pipeline, trained under the same experimental conditions and assessed using the same metrics. The system consists of multi-level feature extraction, texture improvement, global reverse attention, non-local contextual discovery, and hybrid BCE-IoU supervision to improve the boundary localisation and region-level consistency. The experimental results demonstrate that the SINet based on ResNet-101 backbone has the best overall performance with IoU of 0.8564, structure-measure $S_m$ of 0.9120, F-measure $F_\beta$ of 0.9050, enhanced-alignment measure $E_\phi$ of 0.9310 and MAE of 0.0330. The sensitivity study on hyperparameters further demonstrates that the learning rate and weight of BCE–IoU loss can stably converge. The experimental results show the importance of controlled backbone selection for reliable COD performances in military surveillance settings.