Generative adversarial networks for fine-grained retail product recognition

Retail product recognition is the task of automatically identifying and classifying products from images, particularly in retail environments such as supermarket shelves. This problem is considered challenging due to visual clutter, inter-class similarity, and subtle intra-class variations. Accurate recognition of products is crucial for retail automation, inventory management, and consumer applications. In this thesis, we explore the impact of synthetic data augmentation using StyleGAN-XL on fine-grained retail product image classification. Two distinct product categories, snacks and beverages, are considered, each comprising 30 classes with varying numbers of original images. Due to the challenges of data imbalance and limited real-world data, synthetic images were generated via StyleGAN-XL to supplement training data. Four experimental scenarios were developed: Scenario 1, combining 80% synthetic and 20% real images; Scenario 2, trained solely on synthetic images; Scenario 3, trained only on 20% of the original dataset; and Scenario 4, trained exclusively on real images. YOLOv11-CLS was employed for classification, trained for 100 epochs with consistent hyperparameters across all scenarios. Results demonstrate that synthetic data significantly improved model performance, particularly when real data was scarce. In the beverage dataset, several individual classes—especially those with visually distinctive packaging—achieved perfect classification metrics (100% precision, recall, and F1-score), indicating that StyleGAN-XL-generated images can effectively capture fine-grained features. However, this level of performancewas not uniformly observed across all classes, emphasizing the continued relevance of real data and the importance of dataset diversity.