<p>Accurate compliance detection of protective gear worn by electrical workers is a crucial prerequisite for ensuring operational safety and reducing accidents. To address the issues in compliance detection algorithms for electrical scenarios, including the ineffective extraction of multi-scale features of workers’ protective equipment, low feature utilization, and loss of fine details, this paper proposes a compliance detection algorithm for electrical workers’ gear based on TMU-GAN. Firstly, a Multi-Branch Feature Extraction Network (MBFE-Net) is designed, along with a targeted Trident Interactive Attention (TIA) mechanism module. This module effectively extracts feature maps at multiple scales, enhances the interaction of branch-specific information within the three-branch network structure, and significantly improves segmentation accuracy. Secondly, a Multi-Scale Recursive Attentional Feature Fusion module (MSR-AFF) is proposed. This module employs dual attention branches to achieve adaptive feature fusion, while a recursive iteration strategy enhances the fusion capability of skip connections. Based on the above innovative structures, a three-branch segmentation network, MSAF-UNet, with multi-scale adaptive feature fusion capability was constructed. Finally, the concept of Generative Adversarial Networks (GAN) is integrated into a segmentation network, TMU-GAN, which consists of MSAF-UNet and a progressive shrinking discriminator. Experiments conducted on the Tianchi Smart Power Operation dataset from Alibaba Cloud demonstrate that the proposed TMU-GAN algorithm improves mIoU and mPA by 6.31% and 4.72%, respectively, compared to the original UNet network. P-values and F1-scores are also increased by 3.01% and 3.92%, outperforming other classical segmentation algorithms.</p>