This repository contains the implementation of a comprehensive approach to fine-grained image classification applied to three distinct datasets: CUB birds, a combined dataset of CUB birds and FGVC Aircraft, and the FoodX dataset. The goal is to enhance classification performance without significantly increasing the computational requirements of the baseline model, ConvNeXt-V2-Large, by more than 5%. This implementation showcases how strategic adjustments and the incorporation of Selective Channel Recalibration Attention (SCRA) can improve accuracy while adhering to computational constraints.

Fine-grained image classification poses a unique challenge in the field of image analysis due to the high intra-class variance and low inter-class variance. This project leverages a pre-trained baseline model, fine-tuning it for specific tasks across different datasets to assess its adaptability and accuracy. The datasets evaluated include the CUB birds, a combined dataset of CUB birds and FGVC Aircraft, and the FoodX dataset. This work explores various modifications, from data augmentation to architectural adjustments, to improve classification performance.

• CUB Birds: 200 classes.
• Combined Dataset (CUB Birds + FGVC Aircraft): 300 classes.
• FoodX: 251 classes.

Our methodologies yielded significant improvements across all tasks:

• CUB Birds Dataset:

  • Baseline Accuracy: 89.49%
  • Improved Accuracy: 90.40%

• Combined Dataset (CUB Birds + FGVC Aircraft):

  • Baseline Accuracy: 89.48%
  • Improved Accuracy: 89.99%

• FoodX Dataset:

  • Baseline Accuracy: 78.62%
  • Improved Accuracy: 79.44%

These results demonstrate the effectiveness of our approach, showing marked improvements in classification accuracy across all evaluated datasets without substantial increases in computational demands.