An eigenvector approach for obtaining scale and orientation invariant classification in convolutional neural networks

Abstract

The convolution neural networks are well known for their efficiency in detecting and classifying objects once adequately trained. Though they address shift in-variance up to a limit, appreciable rotation and scale in-variances are not guaranteed by many of the existing CNN architectures, making them sensitive towards input image or feature map rotation and scale variations. Many attempts have been made in the past to acquire rotation and scale in-variances in CNNs. In this paper, an efficient approach is proposed for incorporating rotation and scale in-variances in CNN-based classifications, based on eigenvectors and eigenvalues of the image covariance matrix. Without demanding any training data augmentation or CNN architectural change, the proposed method, 'Scale and Orientation Corrected Networks (SOCN)', achieves better rotation and scale-invariant performances. SOCN proposes a scale and orientation correction step for images before baseline CNN training and testing. Being a generalized approach, SOCN can be combined with any baseline CNN to improve its rotational and scale in-variance performances. We demonstrate the proposed approach's scale and orientation invariant classification ability with several real cases ranging from scale and orientation invariant character recognition to orientation invariant image classification, with different suitable baseline architectures. The proposed approach of SOCN, though is simple, outperforms the current state of the art scale and orientation invariant classifiers comparatively with minimal training and testing time. KeywordsRotation in-variance • Scale invariance • Eigenvectors • Eigen values • Principal component • CNN B Deepak Mishra

Disentangled image representation: from affine transforms to facial attributes

“…The attention method in [75] combines the spatial attention mechanism and channel attention mechanism to reduce the spatial variance of the object. Alternatively, the eigenvector approach of [76] applies a scale and orientation correction for images based on eigenvectors and eigenvalues of the image covariance matrix. In adaptive Gabor convolutional networks [77], the convolutional kernels are adaptively multiplied by Gabor filters to achieve invariant information extracted from images.…”

“…The spatial transformer network (STN) [46] is used in [60] to tackle the joint image alignment problem on larger datasets with higher variability. STN [46] In summary, previous approaches for achieving invariance to affine transforms in images have three limitations: i) Some of these algorithms only contain a spatial invariance module [46,47,[74][75][76][77] embedded in a neural network designed for classification, object recognition or other tasks. As this module has to be trained via the learning objectives associated with those learning tasks, it is unable to learn the image transform parameters independently.…”