Abstract:

Objective This study focuses on the critical issue of feature loss in fabric image retrieval under complex real-world conditions, including uneven illumination, fabric folding, and occlusions. These challenges significantly degrade retrieval performance, limiting practical applications. To address this, we propose a novel feature fusion-based method that integrates complementary image features to improve retrieval robustness and accuracy. The research is important for facilitating intelligent fabric identification, which benefits textile production workflows and online fabric commerce.

Method A fabric image dataset with over 28 000 images was constructed, covering diverse complex scenes with varying lighting, folds, and occlusions. Local texture features were extracted using the Scale-Invariant Feature Transform (SIFT), which is robust to scale and rotation. These were aggregated into compact global descriptors via the Vector of Locally Aggregated Descriptors (VLAD). Simultaneously, deep semantic features were obtained from a pretrained Convolutional Neural Network (CNN) model to capture high-level information. Finally, a weighted fusion strategy combined low-level handcrafted and high-level deep features to enhance representation and improve retrieval performance under challenging conditions.

Results Experiments were conducted on a dataset consisting of over 28 000 fabric images collected under diverse and complex conditions. The proposed feature fusion method achieved a mean average precision (P_mAP) of 0.845. The precision at top 5 retrieved images (P) was 78.8%, and the recall at top 5 (R) reached 65.1%. These results reflect high top-ranked relevance and broad retrieval coverage under challenging scenarios. Ablation studies were carried out to investigate the influence of key parameters. When varying the number of VLAD dimensions retained after PCA, retrieval performance changed accordingly, with the best results obtained when 1 024 dimensions were preserved. Different CNN backbone networks were evaluated, including AlexNet, VGG16, ResNet18, and ResNet101. Among them, ResNet101 yielded comparatively better retrieval outcomes. Fusion weight of allocation experiments indicated that assigning higher weights to handcrafted features led to higher P_mAP, P, and R values. Comparisons with representative methods, including handcrafted approaches (MLBP, SIFT-ORB, GCM, FLBP) and deep learning models (VGG16), showed that the fusion method consistently achieved higher retrieval accuracy. Under varying conditions such as illumination changes, fabric folding, and partial occlusion, the method maintained stable performance. Evaluation across different fabric categories and scene complexities demonstrated consistent ranking results. These findings suggest that combining handcrafted and deep features contributes to improved retrieval performance across a range of scenarios.

Conclusion This paper presents a fabric image retrieval method tailored for complex scenarios by fusing handcrafted and deep features. The integration leverages the fine-grained descriptive power of traditional features and the semantic abstraction capabilities of deep learning, resulting in enhanced discriminability and robustness. Experimental results demonstrate that the proposed approach outperforms both conventional techniques and the VGG16 model, achieving a maximum P_mAP 0.875 and P of 78.8% while maintaining efficient retrieval speed. The method exhibits stable performance under challenging conditions such as intricate textures and diverse patterns, making it suitable for practical applications like fabric retrieval and quality inspection. Future work will explore multi-scale feature fusion and lightweight network design to further improve adaptability and scalability across various deployment environments. This study highlights the potential of hybrid feature strategies in bridging the gap between low-level detail and high-level semantics, providing valuable insights for industrial vision systems.

Key words: fabric retrieval, feature fusion, complex scenario image, scale-invariant feature transform, convolutional neural network