Development of IoT Enabled Deep Learning Model for Indian Food Classification: An Approach Based on Differential Evaluation

Due to its extensive use in several areas, deep learning has attracted much interest in the past 10 years. Furthermore, decision-making applications for IoT devices are required, and the number of such devices is growing exponentially. Conversely, IoT devices are subject to resource constraints such as limited power, memory, and computation power. As a result, deep learning models that require less storage space and have a shorter inference time are more popular than traditional models. In the proposed article, we have discussed a differential evaluation-based approach for optimizing the storage space with a significant decrease in inference time without compromising the accuracy too much. We used an openly available Indian food dataset for the experimental work, using popular pre-trained architectures for classification purposes. We then compress the trained models using the differential evaluation approach. The simulation results show that the VGG16 architecture is compressed by 46.15%, with a decrease in precision of 1.91%.