基于多尺度空频动态融合的在架图书错序检测方法

Multi-Scale Spatial-Frequency Fusion Network Based on Wavelet Transform for Misplaced Book Detection on Shelves

摘要:

针对高校图书馆智能化转型过程中在架图书错序管理的难题,提出了一种基于小波变换的空频双域协同检测方法。该方法设计了多尺度空频动态融合模块,通过Haar小波分解实现书脊几何形变特征的下采样提取,并结合逆小波变换对全局纹理分布进行重建,充分发挥空域与频域信息的互补优势,实现多尺度特征的深度融合。为提升分类与定位的协同训练效果,引入匹配感知损失函数以优化特征匹配质量。针对文本区域易受光照与复杂背景干扰的问题,构建了形态学-PMD扩散协同预处理流程,融合自适应伽马校正与动态结构元素闭运算,显著增强了文本区域的鲁棒性。实验结果表明,该方法在书籍密集等复杂场景下,平均精度指标mAP@0.5:0.95、mAP@0.75和mAR@0.5:0.95分别达到了80.2%、96.8%和88.6%,为智能书架管理系统提供了可靠的技术支持。

Abstract:

To address the challenge of managing mis-shelved books during the intelligent transformation of university libraries, this paper proposes a spatial–frequency dual-domain collaborative detection method based on wavelet transform. The method incorporates a multi-scale spatial-frequency dynamic fusion module, in which Haar wavelet decomposition is employed to downsample and extract geometric deformation features of book spines, while inverse wavelet transform is used to reconstruct global texture distribution. This design fully leverages the complementary advantages of spatial and frequency domain information, achieving deep fusion of multi-scale features. To enhance the joint training of classification and localization, a match-aware loss function is introduced to optimize feature matching quality. For text regions susceptible to illumination variations and complex background interference, a morphological–PMD diffusion collaborative preprocessing pipeline is developed, integrating adaptive gamma correction with dynamic structural element closing operations, thereby significantly improving text-region robustness. Experimental results demonstrate that in complex scenarios such as densely arranged books the proposed method achieves average precision metrics of mAP@0.5:0.95, mAP@0.75, and mAR@0.5:0.95 at 80.2%, 96.8%, and 88.6%, respectively, providing reliable technical support for intelligent bookshelf management systems.