ALDS-YOLO: A Lightweight Defect Detection Model for Warehouse Carton Goods

The integrity of carton packaging is critical to ensuring the efficient circulation of goods in large-scale warehouses. Due to the visual obstructions caused by high stacking of goods, uncrewed aerial vehicles (UAVs) are used for aerial visual inspection. However, limited by onboard computational resources, the detection model must be lightweight. Meanwhile, the diversity of printed patterns on cartons leads to complex backgrounds for defect detection and high interclass similarity of defect features, so the difficulty of detection is further aggravated. Therefore, we propose an adaptive and lightweight dynamic sharing-you only look once (ALDS-YOLO) method for visually inspecting large warehouses using UAVs. First, an adaptive and lightweight dynamic (ALD) module is designed to dynamically adjust the weight distribution of convolutional kernels. This effectively enhances the adaptive feature extraction ability of the model and the differentiation ability of the features with high similarity between classes while reducing the computational complexity. Second, a refined lightweight path aggregation network (RLPANet) is designed to strengthen the focus on channel and spatial information in bidirectional feature interaction. This enables the network to capture key defect features under complex backgrounds more accurately. Finally, a lightweight shared convolutional detection (LSCD) is introduced to reduce the computational complexity by using shared convolution with a compensation mechanism, which avoids independent convolutional operations in each layer. Experimental results on the self-constructed warehouse storage defect inspection dataset (WSDID) show that compared with the baseline model YOLOv11n, the proposed ALDS-YOLO reduces the number of parameters by 11.54%, decreases floating-point operations (FLOPs) by 36.51%, and improves detection accuracy by 2.5%, which demonstrates that the proposed method better balances the detection accuracy and speed. In addition, the generalization experiments on three public datasets, Cardboard Defect, Welding Defect, and GC10-DET, further verify the effectiveness and practicality of ALDS-YOLO in industrial scenarios.