Design of Kernels in Convolutional Neural Networks for Image Classification

摘要

Despite the effectiveness of Convolutional Neural Networks (CNNs) for imageclassification, our understanding of the relationship between shape ofconvolution kernels and learned representations is limited. In this work, weexplore and employ the relationship between shape of kernels which defineReceptive Fields (RFs) in CNNs for learning of feature representations andimage classification. For this purpose, we first propose a featurevisualization method for visualization of pixel-wise classification score mapsof learned features. Motivated by our experimental results, and observationsreported in the literature for modeling of visual systems, we propose a noveldesign of shape of kernels for learning of representations in CNNs. In theexperimental results, we achieved a state-of-the-art classification performancecompared to a base CNN model [28] by reducing the number of parameters andcomputational time of the model using the ILSVRC-2012 dataset [24]. Theproposed models also outperform the state-of-the-art models employed on theCIFAR-10/100 datasets [12] for image classification. Additionally, we analyzedthe robustness of the proposed method to occlusion for classification ofpartially occluded images compared with the state-of-the-art methods. Ourresults indicate the effectiveness of the proposed approach. The code isavailable in github.com/minogame/caffe-qhconv.