PDNet: a lightweight attention-guided CNN for efficient pallet racking defect detection on edge devices

Abstract This study presents three algorithmic innovations aimed at optimizing Convolutional Neural Network architectures for automated defect detection in pallet rack inspection scenarios. First, a domain-specific Augmentation Algorithm is introduced to expand the training dataset, mitigate overfitting, and enhance the model’s generalization by emphasizing critical visual features associated with structural defects. Second, a guided CNN Development Mechanism facilitates architectural optimization by systematically refining filter sizes, neuron counts, and convolutional block configurations, enabling high performance with reduced parameter overhead. Third, the proposed PalletDetect Module (PD-M) enhances computational efficiency by adaptively refining feature representations at the input tensor level, reducing complexity while preserving discriminative capacity. These algorithms collectively produce PDNet, a compact CNN that enables real-time pallet racking inspection on resource-constrained edge devices. PDNet achieves an accuracy of 92.07%, with a computational complexity of only 32.31 million multiply–accumulate operations (MMAC) and a compact memory footprint of 31.36 MB. Compared to modern lightweight CNNs such as MobileNetV3 and ShuffleNetV2, PDNet offers a superior balance between accuracy, speed, and computational efficiency, demonstrating its potential for real-time industrial inspection applications.