An efficient learning-based multi-scale method for Food image classification by Jiran Ou Master of Applied Statistics & Data Science University of California, Los Angeles, Food image classification has become an important task within the field of computer vision, supporting applications such as dietary monitoring, automatic food recognition, and nutritional analysis. Earlier methods primarily relied on handcrafted features and shallow architectures, which often struggled with the diverse visual appearances of food items. With the advancement of deep learning, particularly the development of convolutional neural networks (CNNs) and vision transformers (ViTs), classification performance has seen notable improvements. Nevertheless, issues such as significant intra-class variation, inter-class similarity, and limited availability of annotated datasets continue to pose challenges. In this study, we introduce a novel deep learning framework tailored for food classification, aiming to enhance feature extraction and improve recognition accuracy. Our approach replaces standard convolutional operations with depthwise separable convolutions, thereby reducing computational overhead without compromising representational power. Furthermore, we incorporate multi-scale feature learning by employing 5×5 and 7×7 convolutional kernels, enabling the network to capture spatial patterns across a range of receptive fields. We validate our method on the Food-11 dataset, a widely recognized benchmark for food classification tasks. Experimental results demonstrate that our approach achieves superior performance compared to existing techniques while maintaining high computational efficiency. These observations highlight that the integration of efficient convolutional operations and multiscale learning strategies can significantly enhance the capability of food image classifiers, particularly in real-world deployment scenarios.