Enhanced EfficientNet-B0 with Dual Attention Mechanisms for Food Category Classification in X-ray Images

To improve the operational efficiency of food foreign object inspection systems, this study proposes a food category recognition model for X-ray images. 6800 X-ray images containing 34 types of food were collected. Image labels were automatically assigned according to the folder corresponding to each food category. The dataset was divided into training, validation, and testing sets in an 8:1:1 ratio. The model was based on EfficientNet-B0. To enhance the representation capability of key image regions, Convolutional Block Attention Module (CBAM) and Efficient Channel Attention Module (ECA) have been added to EfficientNet-B0. To improve the robustness of the model to noise and image variability, the images were enhanced through various methods such as random scaling, horizontal flipping, rotation, occlusion and so on. The proposed model achieved a test accuracy of 96.60%, which is 2.03% higher than ResNet-50, 1.56% higher than the baseline EfficientNet-B0, and 2.14% higher than MobileNet-V2. The test results indicate that the proposed method can provide accurate, efficient, and automated technical solutions for food classification in X-ray images.