Invariant feature extraction for image classification via multi-channel convolutional neural network

Abstract

The invariance for feature extraction, such as invariance for specificity of homogeneous sample and rotation invariance, is crucial for object detection and classification applications. Current researches mainly focus on a specific invariance of features, such as rotation invariance. In this paper, a novel multi-channel convolutional neural network (mCNN) is proposed to extract invariant features for object classification. Multi-channel convolutions sharing identical weights are used to alleviate the feature variance of sample pairs with different rotations in the same category. As a result, the invariance for specificity of homogeneous object and rotation invariance are simultaneously encountered to improve the invariance of features. More importantly, the proposed mCNN is especially effective for small training samples. Experimental results on two benchmark datasets for handwriting recognition demonstrate that the proposed mCNN is very effective to extract invariant feature with small amount of training samples.