Comparative Analysis of Neural Architecture Search Methods for Classification of Cultural Heritage Sites

Abstract

AbstractIn the current era of Machine Learning, the performance of Neural Networks in object detection, image classification, and video analytics has improved with better design of architecture. It often requires engineers to spend substantial time and effort to design the network, which can be an error-prone method. This method does not exhaustively search the entire search space of possible neural network architecture and guarantees optimal accuracy from the designed model. Neural Architecture Search (NAS) automates this process to find the optimal network to outperform the hand-designed model. Though NAS methods have shown promising performance for image classification tasks, it is challenging to infer why they work well on standard data sets and perform poorly when transferring the same NAS method to custom/real-time data sets. This paper proposes a custom image data set based on Indian Heritage sites built using a crowdsourced framework to perform a comparative performance analysis of NAS methods for the image classification task. The data set consists of 20,000 color images of 1920*1080 pixels from 40 heritage sites, with 16,000 training and 4000 test images. The comparative study is performed on three primary NAS methods viz. Efficient Neural Architecture Search via parameter sharing (ENAS), Differentiable Architecture Search (DARTS), and Neural Architecture Search using Multi-Objective Genetic Algorithm (NSGA-Net). The DARTS showed 88.625% accuracy, ENAS showed 32.83% accuracy and NSGA-Net produced 69.92% accuracy on the custom data set.KeywordsDeep learningImage classificationNeural architecture searchCultural heritage sitesGradient-based optimizationEvolutionary algorithmReinforcement learning