An Edge Distortion and CNN-Based Analysis of Blind IQ

Abstract

This paper is for assessing the image quality (IQ) without using an authentic image (original image) which is a type of Blind IQ Assessment (BIQA) model by introducing a technique of Convolutional Neural Network (CNN). The distortions of edges in the image are considered as features to represent the image feature vector. This approach is justified by the evidence that the subjective evaluation concentrates on image characteristics that radiate from the boundaries and edges that exist with in the image. It was identified in the prior methods that the features are extracted at the time of training or before training by applying sophisticated transformations on the image. In this work, the vertical along with horizontal edge feature maps of the training images are extracted by means of Scharr Kernel (SK) approach. These edge maps subsequently fed into a CNN, which uses non-linear transformations to bring out higher-level features. Regression is then used to link the generated features to the IQ score. To accommodate different sizes of input images, the SPP (Spatial Pyramid Pooling) layer is used in this network. The developed model was evaluated using well-known datasets in the field of IQA. The suggested model's performance reveals that it outperforms previously existing models in context of negligible complexity involvement and feature extraction simplicity.