A Modified Transfer Learning-Based Framework for Efficient Food Image Classification

Accurate food classification of images has become a major concern in the age of digital food sharing, with far-reaching implications for nutritional monitoring, dietary analysis, and health-related applications. In this study, we conducted overfitting experiments in deep learning models for food image classification. Similarly, conventional methods obtain superior training accuracy however fail in the generalization making it non-reliable mode of working. We have tested EfficientNetB7, Resnet50 and VGG19 with the Food-11 dataset in this study, which gives us results that when applied to our task of interest are best performed by VGG19 but overfitting was seen. To improve regularization, the structure (fully connected layers) was changed to VGG19 Variants, enhancing generalization and accuracy while providing insights for developing robust deep learning models for food image classification. EfficientNetB7 obtained 94.13% training accuracy and 87.38% validation accuracy, whereas ResNet50 achieved 85.85% and 80.87%, respectively. The variants of VGG19, such as VGG19 with Frozen Convolution Layers and a Single 512-Unit MLP Layer (VGG19-512) obtained 68% accuracy, VGG19 with 50% Trainability and Multiple MLP Layers with Dropout (VGG19-512-128-50p-T) reached 79% accuracy, VGG19 with Trainable and Multiple MLP Layers with Augmentation (VGG19-512-128-T-Aug), and VGG19 with Partial Trainable and Multiple MLP Layers with Augmentation, Dropout, and EarlyStop (VGG19-GAP-512-5-128-3-T-Aug-EarlyStop) got the same-83% accuracy, respectively. This illustrates a need for further investigation in novel model structures as well as regularization techniques written over deep learning-based image classification mechanisms.