Abstract
Inspection stations on logistics sorting lines image metallic components with photoelectric cameras on a moving conveyor, and the images are degraded by belt-motion blur, uneven illumination, sensor noise, and compression. We propose LSDD-Net, a lightweight convolutional network that couples an illumination-robust edge-enhancement (IREE) stem, a hybrid coordinate–channel attention (HCCA) module, and a multi-scale feature (MSF) fusion head. The study addresses the classification of pre-localised defect regions. It uses the six-class NEU steel-surface dataset together with a simulated acquisition–degradation model, because real conveyor images were not available. On a fixed degraded test split, LSDD-Net reached 98.61% accuracy and 98.62% macro- with 0.43 M parameters and 0.06 GFLOPs. Over five-fold cross-validation repeated three times, it obtained % degraded accuracy. This was significantly higher than ResNet-18/50, EfficientNet-B0, VGG-16, ShuffleNetV2, and a HOG–SVM pipeline (McNemar ) and statistically on par with MobileNetV2 ( %, image-level McNemar ), which has five times as many parameters. Single-factor tests on clean-trained models showed that the edge stem raises accuracy by 4.5 and 5.5 points under strong global illumination change and non-uniform shading, but lowers it under sensor noise. The model ran in 0.42 ms per image with TensorRT FP16 on an A800 GPU and in 1.9 ms on four CPU threads. No real conveyor images or embedded-device measurements were available, so these results show robustness under simulated degradation, not field performance.