Advanced Image Classification on Intel Datasets Using Optimized EfficientNet and MobileNetV2

In this paper, an extensive analysis is provided to improve EfficientNet and MobileNetV2, two known neural network architectures commonly used in advanced image classification based on Intel image datasets. The machine learning application is rapidly growing and there has never been a time that the companies need high-accuracy, computationally efficient image classification models. Such ahead-of-their-time demand is represented by EfficientNet and MobileNetV2 each providing the best possible efficiency with an accuracy/efficiency tradeoff. Yet, the capability of adjustment under particular Dataset challenges such as would be met by heterogeneous and sophisticated Intel image data is just moderately developed. This study saw us dive deeper into the specific optimization methods for such architectures, which we aspired to improve Intel dataset performance. Lying at the basis of our methodology were several strategic modifications such as scaling strategies to implement EfficientNet and angular tuning for MobileNetV2, along with automatic mixed precision training which posed universal applicability against neural architecture search. The outcome of our optimizations showed that the model accuracy improved drastically and the boosted EfficientNet scored an overall accuracy of 94.5% while MobileNetV2 gave a score of 92%, thus, huge improvements from their base configurations were observed. Additionally, the pruning in our study went deep into the computational efficiency of these models where we observed a significant decrease in both inference and model size while maintaining the same accuracy. These results not only support the performance of the proposed optimizations but also emphasize upon technique utilized for problem-specific solutions.