LFRSCNet: Skin defect detection based on lightweight flexible residual separable convolutional network

Aircraft skin is prone to surface damage, such as cracks and dents, during long-term service or manufacturing processes. These defects not only threaten structural integrity but may also pose potential safety hazards. The industrial sector continually explores more efficient and precise detection methods to address this issue. Therefore, this paper proposes a skin defect detection method based on a lightweight and flexible residual separation convolutional network to improve detection accuracy and efficiency. Therefore, this paper proposes a skin defect detection method based on a lightweight flexible residual separable convolution network to improve detection accuracy and efficiency. First, a lightweight flexible residual separable convolution module (LFRCM) is designed, which effectively integrates multi-modal features by combining multi-scale receptive fields with an adaptive channel attention mechanism; at the same time, a lightweight backbone network based on PP-LCNet is constructed, employing a collaborative optimization strategy of depthwise separable convolutions and the h-swish activation function to significantly enhance inference speed while maintaining detection accuracy; finally, the MPDIoU metric criterion is introduced, which effectively improves target localization accuracy by implementing a center point offset penalty mechanism. Experiments on the self-built professional dataset SD-DET and the public dataset GC10-DET show that the model achieves [email protected] of 99.5% and 86.2%, respectively, demonstrating significant advantages over mainstream detection models. Systematic ablation experiments confirm the synergistic effect of various innovative modules. Finally, verification experiments are conducted on the AIRCRAFT skin defect dataset, achieving an [email protected] of 30.7%. Quantitative analysis and comparative experiments verify that LFRSCNet can achieve detection accuracy breakthroughs while maintaining low parameter counts and computational costs. Its balanced accuracy-efficiency characteristics provide an efficient and reliable solution for surface defect detection in industrial scenarios.