A Lightweight Book Defacement Recognition Method for Self-Service Book Borrowing and Returning Machines

作者：Liyang Yu, Huaying Liu, Ruilin Deng · 年份：2025 · DOI：10.1109/ictai66417.2025.00124 · 研究领域：Big Data and Digital Economy、Advanced Neural Network Applications、Handwritten Text Recognition Techniques

Self-service kiosks have become the main channel for book circulation in libraries. However, their current processes generally lack effective capability for detecting book damage. Simultaneously, the limited memory resources of these kiosks restrict the integration of complex service components. To address this, this study proposes a lightweight damage classification model named GCF-MobileNetV3. Based on the MobileNetV3 baseline network, the model significantly enhances feature representation capability and computational efficiency by introducing an enhanced Ghost module that integrates deformable convolution and channel shuffle. It innovatively constructs a parallel fusion structure (CA-ECA) combining Coordinate Attention (CA) and Efficient Channel Attention (ECA) modules to collaboratively capture spatial location information of damaged regions and inter-channel dependencies. Furthermore, it replaces the standard ReLU activation function with Funnel ReLU (FReLU), effectively mitigating the issue of neuron deactivation. Experimental results demonstrate that while maintaining a high degree of lightweight design ($\mathbf{5. 2 M}$parameters,$\mathbf{1 9 5 M}$FLOPs computational load), the model achieves a 3.7 % reduction in model size and an 11 % reduction in computational complexity compared to the baseline model. It attains a high classification accuracy of 98.45 %, and effectively adapting to the low-memory constraints of self-service kiosks.