Oil leakage in industrial equipment poses significant risks to safety and operational continuity. Its visual manifestations range from large-scale fluid accumulations on the ground to smaller, elongated drips. Traditional heavy-weight segmentation networks struggle with deployment under the dual constraints of complex industrial backgrounds and real-time requirements, while lightweight methods often lack sensitivity to small targets due to insufficient contextual capture and coarse feature fusion. To address this, we propose a lightweight real-time semantic segmentation method tailored for oil leakage detection. It significantly enhances the perception of small targets and fine boundaries while maintaining high-speed inference.The proposed method first employs a lightweight Atrous Spatial Pyramid Pooling (ASPP) module to extract global features, effectively integrating both long-range contextual information and local details. Subsequently, during the fusion of high- and low-resolution features, a Residual Channel Attention Fusion module is introduced. This module enhances semantic detail and strengthens feature extraction capabilities without adding critical path delays. Finally, the segmentation result is output through a classification module.Experimental results on an industrial oil leakage dataset show that our algorithm achieves an mIoU of 78.56%, which is 4.86 percentage points higher than the original Fast-SCNN. With an inference speed of 112 FPS, it meets real-time detection requirements. This approach effectively mitigates the common weakness of lightweight networks—insensitivity to small targets—offering a practical solution for rapid deployment in industrial scenarios.