Abstract
Surface defects on lithium-ion battery electrode sheets are small and diverse and can affect battery performance and safety. This study addresses two defect types in production-line images: printing defects, which have low contrast and blurred boundaries, and streak defects, which are thin, elongated, and strongly oriented. We propose DMD-YOLO (Direction-Aware, Multi-Receptive-Field and Dynamic YOLO), which adapts existing network components to these defect characteristics on the basis of YOLOv11n. In the backbone, a multi-directional entropy-aware (MDE) module adds four orientation-constrained convolution branches whose fused output is scaled by an image-level coefficient derived from directional entropy. In the neck, a Dual-Branch Gated Dilation Block (DGD-Block), adapted from the dilation-wise residual module, fuses a local and a dilated branch through non-competitive spatial gating and is used to build GD-PANet. In the head, omni-dimensional dynamic convolution is applied to the box-regression branch (ODDetect), and the box loss combines Inner-IoU with a simplified Shape-IoU penalty. We evaluated the method on 2364 production-line images annotated with three classes (electrode-sheet body, printing, and streak). Over three random seeds, DMD-YOLO improved Recall, mAP@0.5, mAP@0.5:0.95, and F1-score over that of YOLOv11n by 2.0 ± 0.7, 1.3 ± 0.7, 2.6 ± 1.0, and 2.2 ± 1.3 points, respectively, with positive differences for every seed, whereas the Precision gain (2.5 ± 3.5 points) varied between seeds. Among the ten lightweight detectors compared (at most 4.23 M parameters), it achieved the highest Precision, mAP@0.5, mAP@0.5:0.95, and F1-score. It did not achieve the best accuracy overall: RT-DETR-R18 and LSRF-YOLO reached higher Recall, mAP@0.5, mAP@0.5:0.95 and F1-score values, although with four to five times as many parameters, and DMD-YOLO requires more computation and runs slower than YOLOv11n (121.5 vs. 208 FPS for model inference). When retrained on the GC10-DET steel surface defect dataset, DMD-YOLO improved Precision and mAP@0.5 over YOLOv11n.