Learning spatial hierarchies of high-level features in deep neural network

Abstract

This paper addresses a new approach to learn perceptual grouping of the extracted features of the convolutional neural network (CNN) to represent the structure contained in the image. In CNN, the spatial hierarchies between the high-level features are not taken into account. To do so, the perceptual grouping of features is utilized. To consider the intra-relationship between feature maps, modified Guided Co-occurrence Block (mGCoB) is proposed. This block preserves the joint co-occurrence of two features in the spatial domain and it prevents the co-adaptation. Also, to preserve the interrelationship in each feature map, the principle of common region grouping is utilized which states that the features which are located in the same feature map tend to be grouped together. To consider it, an MFC block is proposed. To evaluate the proposed approach, it is applied to some known semantic segmentation and image classification datasets that achieve superior performance.