Multi-Scale Localization and Attention Mechanism for Fine-Grained Image Classification

Fine-grained image classification tasks present a notable challenge, images within the same class can exhibit substantial intra-class variation, while the differences between distinct classes are often minimal. As a result, identifying the key distinguishing features in the images is essential for tackling this issue. This paper introduces an Attention Object Localization Module (AOLM) that predicts object locations, combined with a Weakly Supervised Part Localization Module (WPLM) to identify informative parts of the image. Additionally, a Convolutional Block Attention Module (CBAM) is applied to highlight important features and suppress less relevant ones. To enhance the model's classification accuracy, a multi-class ResNet backbone is used. Evaluations on three well-known benchmark datasets—CUB-200-2011, Stanford Cars, and FGVC-Aircraft—resulted in accuracies of 90.5%, 95.0%, and 95.0%, respectively, showcasing the effectiveness of the proposed method in fine-grained image classification.