Morphological Difference of Closings Operator for No-Reference Quality Evaluation of DIBR-Synthesized Images

Abstract

Images synthesized using Depth-Image-Based Rendering (DIBR) techniques are characterized by complex structural distortion. Multi-resolution multi-scale sparse image representation generated using morphological Difference of Closings operator (DoC) is used to efficiently capture structure-related distortion of synthesized images in the noreference DoC-GRNN image quality assessment model. Nonlinear morphological Difference of Closings operator (DoC) with an array of line-shaped structuring elements of increasing length is used to extract perceptually important details of object structure at different scales and resolutions. The sparsity of DoC band is calculated as scalar feature. The extracted features are mapped to the quality score by general regression neural network (GRNN). We have explored the influence of the direction of an array of line-shaped structuring elements on the model's performances. The DoC-GRNN model shows high agreement with perceptual quality scores, comparable to the state-of-the-art metrics, when evaluated on the stereoscopic DIBR-synthesized images of MCL-3D dataset.

No-reference quality assessment for depth image based rendering-synthesized images based on multi-feature fusion

“…The method directly learns and predicts the quality of DIBR-synthesized images, overcoming the limitations of traditional methods. Sandic ´-Stankovic ´et al 28 proposed a quality model, called the difference of closings operator-general regression neural network model, to measure the structuredependent distortion of DIBR-synthesized images without references. Sadbhawna et al 29 detected distorted regions of synthesized images and evaluated the quality of the image by using a convolutional neural network.…”