Enhanced Perceptual Feature Extraction for Blind Image Quality Assessment using Extreme Learning Machine

Abstract

In recent years, blind image quality assessment (BIQA) has become a major research topic with the practical applications as it promises the effective outcome as compared to the others image quality metrics. In situation where the reference image is unavailable, BIQA is the most suitable approach. However, the state- of-the-art image quality assessment (IQA) metrics is specific to certain types of distortion and not consistent with human perception. These existing BIQA do not consider for human visual characteristics impact on image content and specific to certain types of distortion. To overcome the problem, the research study proposed perceptual based features during initial data collection and combine with pooling algorithm based on Extreme Learning Machine (ELM). This is to match human perception that can see noticeable difference at certain frequency range and conclude the value similar to neuron. The approach aims to develop the lifting wavelet-based feature extraction with the aid of extreme learning machine (ELM) algorithm that able to enhance the image quality, extract the significant features from the image and reduce noise to the maximum extent possible. Unlike discrete wavelet transform (DWT), lifting wavelet transform (LWT) provides low computational complexity as it does not require convolution, dilation, and translation of the traditional mother wavelets. Besides, the pooling strategy of ELM is employed to overcome the limitations of previous pooling techniques such as neural networks (NNs) and support vector regression (SVR). This is because ELM has higher learning accuracy with faster learning speed. The proposed approach is verified on several type of image database that consist of different distortion type to make it general based BIQA. Based on experimental result, it proved that the approach has good performance than other features in terms of accuracy, specificity, and sensitivity. The outcome of this work will be essential in various image processing applications such as optimizations of image enhancement in medical for tumour or cancer detection, image watermarking for security detection, image coding and compression or image forensic. In biodiversity monitoring, image enhancement plays an important role for tracking and data collection.