Feature Sampling based on Multilayer Perceptive Neural Network for image quality assessment

Abstract

Image quality assessment (IQA) has a vital issue in image processing to measure the perceptual quality of the image. This aims at the human visual system (HVS) by viewing for distorted images from reference images. HVS is the ultimate receiver of visual data in major applications. But, quality assessment is generally a time-consuming and luxurious method. In our work, the novel Bilateral Smoothing Filtered Feature Sampling-Based Tanimoto Index’ve Multilayer Perceptive Neural Network (BSFFS-TIMPNN) technique is developed. The main contribution of the designed BSFFS-TIMPNN technique is for enhancing the quality of the assessment performance with full reference and no reference images with multiple layers. At first, the original and reference image is taken to the input layer and sent to the first hidden layer. Followed by, image preprocessing is performed in the first hidden layer by using the novelty of a nonlinear bilateral smoothing filtering technique for eradicating noisy pixels with a higher-quality image. Next, patch extraction is executed in the second hidden layer with the novelty of the nearest neighbor sampling approach to split images into various patches. Then, the different features of the image such as shape, color, and texture are extracted in the third hidden layer. Finally, classification is carried out output layer by applying the innovation of the Tanimoto similarity index to examine the extracted features for IQA. We estimate the proposed BSFFS-TIMPNN technique on the Tampere Image Database (TID2013) dataset with qualitative and quantitative results analysis. The proposed BSFFS-TIMPNN technique is employed to compute image quality with maximum accuracy and minimum time and memory consumption when compared to other related methods by MATLAB simulator. The observed quantitative result demonstrates the better performance of the proposed BSFFS-TIMPNN technique by full reference and no reference with higher PSNR by 8%, accuracy by 15%, 17%, lesser memory consumption by 15%, 17, and faster prediction time by 9%, 9% than the state of art methods respectively.