Optimization of CNN model for image classification

Abstract

Convolutional Neural Networks (CNNs) are particularly precise in several fields, especially computer vision where image classification is one of the most researched and commercialized application. Deploying these models on embedded devices requires high throughput and low latency even with limited resources and energy budgets. The complexity of the architecture of CNN models implies a very high computation cost. We are looking in this paper for determining the optimal topology (the number of layers and the number of neurons per layer) that allows us to reduce the model and deploy it in embedded platforms. We have proposed a small CNN architecture that achieves high level accuracy 81.50% on CIFAR-10 with fewer parameters based on the growing approach. For more optimization we used the pruning technique and the results showed that with more optimal architecture we obtained 82.43% accuracy and 15% reduction of the number of parameters.