DAMR-GMBINet: A Robust Multi-Scale Lightweight Segmentation Network for Micro-Defect Detection in Complex Degraded Environments
Abstract
In precision manufacturing driven by Industry 4.0, pixel-level segmentation of surface defects is crucial for accurate quality control. However, existing lightweight segmentation networks generally face two major bottlenecks: fixed receptive field design struggles to balance noise suppression and detail preservation, and high-level semantic responses tend to overwhelm weak micro-defect signals during cross-scale feature fusion, leading to a significant decline in small-target segmentation accuracy. Moreover, the robustness of models under complex degraded environments such as noise and low illumination remains to be improved. To address these issues, this paper proposes a degradation-aware multi-scale robust lightweight segmentation network (DAMR-GMBINet) based on the GMBINet framework. On the encoder side, a degradation-aware dynamic scale interactor module is designed, where a degradation prior encoder drives adaptive interpolation of dilation rates for atrous convolutions and constructs a bidirectional gating mechanism with forward detail guidance and backward semantic correction to achieve dynamic receptive field adjustment. On the decoder side, a micro-defect fidelity interactor module is proposed, which decouples feature fusion into two branches: main structure modeling and high-frequency detail compensation, and introduces an area-aware enhancement mechanism to adaptively strengthen high-frequency responses of small targets according to defect scales. During training, a robust consistency regularization protocol is built, imposing dual-view consistency constraints on multiple levels such as prediction outputs, intermediate features, and degradation priors, driving the network to learn degradation-insensitive representations while keeping zero extra overhead during inference. Experimental results show that DAMR-GMBINet achieves competitive clean-scene accuracy while maintaining the highest robustness across a wide range of degradation conditions: it improves the average foreground IoU over 12 degradation conditions (IDR-12) by 14.76 percentage points over the best competing lightweight network, with the gains further supported by statistical significance analysis. The proposed method provides an efficient and robust technical solution for online precision quality inspection in complex industrial environments.