Research on fine-grained retrieval based on deep learning and fusion loss algorithm

Fine-grained image retrieval (FGIR) has emerged as a popular research topic in recent years. Researchers aspire to design more intelligent and efficient models to drive the development of fine-grained image retrieval. To tackle the issues that fine-grained image retrieval demands annotation and location assistance and that accurately retrieving similar images is complex, researchers aim to enhance the expression ability of convolutional neural networks for feature extraction and boost the robustness of models against noise and outliers. In this paper, we discover that lightweight attention mechanisms enhance feature representation by aggregating spatial and channel attention information within the model. Additionally, fusion loss functions can improve the robustness to data noise and outliers through multiple constraints in the model network. We propose a feature extraction and aggregation method (FALF) that combines the attention mechanism with the fusion loss function. Firstly, VGG16Net is selected as the backbone network of the model, and the CBAM module is added after the C1, C2, C3, C4, and C5 modules. The weighted feature map obtained by fusing and multiplying CBAM's unique spatial and channel attention is added to the original feature map to strengthen feature representation. Secondly, the fusion loss function replaces the original loss function of the model. The optimizer parameters are adjusted, the model is retrained, and the decision process and performance evaluation results of changing the loss function are recorded. Compared with previous enhancements, the model shows an improvement of 6.56% and 10.98% on the CUB-200-2011 dataset and a self-built dataset in terms of mAP@1, respectively.