Because of variations in light, angle, size, occlusion, and noise in the background, image recognition remains one of the most difficult problems in vision. This is due to the ability of Convolutional Neural Networks (CNNs) to learn hierarchical feature representations without human intervention in a nonlinear manner through local receptive fields, weight sharing, and spatial subsampling and therefore have become the state-of-the-art method for this problem. We introduce an improved CNN architecture that uses smaller convolutional kernels, organized in deeper stacks and regularized during training to yield higher accuracy than previously used architectures, while consuming less CPU resources. The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) benchmark dataset is used to evaluate the results of the proposed model in comparison to state-of-the-art methods. Experimental results showcase a Top-5 error of 9.18%, outperforming multiple state-of-the-art methods whilst sustaining high scalability properties and low training costs.
Kumar et al. (Wed,) studied this question.