WHEAT-Net: Real-time Food Object Detection Network Based on Efficient Feature Extraction and Fusion

Tracking food diet is crucial for effective health control. The rapid advancement of portable equipment has enabled automated food recognition through images, creating food classification a famous study region in artificial intelligence. However, problems persist, such as the limited access to public datasets, high mathematical costs, and poor accuracy for deployment on portable devices. Improvement of detection accuracy and shorter model size are necessary for the development of food recognition systems for health monitoring. This paper explores the development and optimization of a five-category food classification and detection model based on YOLOv8. This paper first integrated the WTConv module into the original C2f module within the backbone of YOLOv8, resulting in the creation of the novel C2f-WTConv module. Furthermore, the paper introduced the Hierarchical Scale-based Feature Pyramid Network (HS-FPN) and the Efficient Local Attention (ELA) module, further improving the model's performance. In response to the challenges in food detection, this study addresses several key issues and presents the following contributions: a 5-class food dataset (comprising staples, vegetables, fruits, meat and soups) with 10000 samples; experiments conducted using this dataset and several mainstream neural networks; and an improved YOLOv8 model called WHEAT-Net that outperforms multiple benchmark models and state-of-the-art methods. The proposed model achieved mAP@50 and mAP@50-95 scores of 0.937 and 0.623, respectively.