Research and Application of Fine-Grained Image Classification Based on Small Collar Dataset

Area of Science:

• Computer Science
• Artificial Intelligence
• E-commerce Technology

Background:

• The growth of apparel e-commerce necessitates accurate classification of garments based on design features like collars.
• Traditional image processing methods struggle with complex backgrounds in garment images, hindering effective classification.
• Fine-grained classification of apparel, particularly collar designs, is challenging due to subtle visual differences.

Purpose of the Study:

• To propose and evaluate an improved classification algorithm for garment collar images.
• To address the limitations of traditional methods in handling complex backgrounds and achieving high classification accuracy.
• To enhance the feature extraction capabilities for fine-grained apparel classification tasks.

Main Methods:

• Development of the EMRes-50 classification algorithm, integrating the ECA-ResNet50 model.
• Incorporation of the MC-Loss loss function method to improve classification performance.
• Experimental validation on the Coller-6 and DeepFashion-6 datasets.

Main Results:

• The EMRes-50 algorithm achieved 73.6% accuracy on the Coller-6 dataset.
• The algorithm demonstrated 86.09% accuracy when applied to the DeepFashion-6 dataset.
• The proposed model outperformed existing Convolutional Neural Network (CNN) models in accuracy and feature extraction.

Conclusions:

• The developed EMRes-50 algorithm offers superior performance for garment collar image classification compared to existing CNN models.
• The enhanced feature extraction ability of the model is crucial for tackling the complexities of fine-grained collar classification.
• This advancement contributes to improved clothing product image classification, benefiting the apparel e-commerce sector.