Multi-layer Cross-Scale Coupling Feature Pyramid Network for Food Logo Detection

Food logos typically serve as a visual representation of a food brand or product by using unique and recognizable images or graphics that are related to the brand or product, such as food items, utensils, or cooking equipment. The study of food logos has important real-world applications, such as in food safety, self-service shops, food recommendation systems, and food brand management. Although detection technology is developing rapidly, these techniques generally suffer from the issue of small logos. Small food logos usually have fewer pixels, making it difficult to extract discriminative features. To address this problem, a Multi-layer Cross-Scale Coupling Feature Pyramid Network (MCCFP-Net) is proposed, which can enhance semantic information by coupling multi-layer features. Specifically, the cross-scale connections break through the limitations imposed by neighboring nodes, enabling the learning of more fundamental features and higher-level semantics. Moreover, to address the problem of similar appearance of different logo categories in classification tasks, the Feature Transformation Offset and Weight (FTOW) is designed to adaptively change the feature regions and assign feature weights. Additionally, Side-Aware Boundary Localization (SABL) is adopted to locate the small food logo objects more precisely. Extensive experimental results demonstrate the effectiveness of our proposed MCCFP-Net.