Spprnet: A Robust CNN for Library Material Recognition via Spatial Pyramid Pooling and Heterogeneous Convolution

In library environments, the diverse scales of materials and batch instability caused by inconsistent scanning conditions pose challenges for image recognition tasks. Traditional ResNet architectures, due to their fixed input size constraints, may reduce their recognition accuracy for images of arbitrary sizes. In this study, we introduce a novel heterogeneous convolution strategy, adjust batch normalization operations, and incorporate a spatial pyramid pooling module based on the ResNet18 network to eliminate these limitations. This new architecture, termed SPPRNet, supports flexible processing of arbitrary-sized inputs and combines multi-scale convolution kernels (3 $\times$ 3, 5 $\times$ 5, 7 $\times$ 7) to simultaneously capture fine-grained features and global contextual patterns. Quantitative results on general datasets demonstrate that our method achieves a 25.47% Top-1 error rate on ImageNet (compared to 30.55% for ResNet18) and attains 92.95% mAP on the Caltech-101 dataset for object detection tasks, outperforming mainstream models such as VGG-16 and MobileNet. The robust performance of our method in image tasks can be extended to existing approaches to further improve the quality of image recognition in library scenarios.