Date of Award

28-3-2024

Document Type

Thesis

School

School of Electrical & Electroncis Engineering

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.N.R.Raajan

Keywords

SLKOF, CGIHE-VDSR, Super Resolution, Denoising Optical Flow, Tetrolet Transform

Abstract

Motion is a significant component of the modern visual experience, aiding with oculomotor control, object recognition, scene interpretation, perceptual organization and the acquisition of 3D shapes. Identifying motion vectors that characterize two-dimensional image transitions, usually from video frames, is necessary to estimate motion.

Optical flow is a motion-tracking technique that estimates the trajectory and velocity of a moving object in a video by shifting each pixel between frames using displacement vectors. This work has two objectives based on small and large displacement optical flow and computes optical flow using Lucas-Kanade and Brox flow, which seem to be superior to other methods.

A novel Subsampled Lucas-Kanade Optical Flow (SLKOF) algorithm has been proposed to detect Opto Kinetic Nystagmus (OKN) by analyzing the human eye's small displacement. Adults experiencing limited visual perception may have neurological issues requiring OKN detection. Images are subsampled in the SLKOF, and the OKN gain is assessed over various subsampling factors.

The OKN experiment involves computer-controlled drum rotation via a stepper motor. The proposed method is compared with Lucas-Kanade (LK) optical flow based on OKN gain. The comparison illustrates that OKN gain of ¼ subsampling factor of SLKOF correlates with LK optical flow for 80% of cases. SLKOF algorithm for OKN detection with reduced image storage size and computation time has been developed.

The visual image quality of images captured by modern cameras is degraded by noise. By combining Contrast Limited Adaptive Histogram Equalization (CLAHE) and Denoising Convolutional Neural Network (DnCNN), the proposed CLAHE - DnCNN algorithm can denoise color images. The CLAHE - DnCNN algorithm is compared to Color Block Matching 3D-based adaptive TV denoising (CBM3D) and Color Denoising Convolutional Neural Networks -Blind (CDnCNN-B) methods using Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics. On average, the proposed algorithm outperforms all test images.

The small size sensor in the Charge Coupled Device (CCD) camera reduces image resolution. Color Global Image Histogram Equalization (CGIHE) is integrated with Very Deep Super Resolution (VDSR) Network in the proposed CGIHE -VDSR algorithm to improve color image resolution. The PSNR and SSIM metrics reveal that the proposal surpasses the benchmark methods.

The Brox and SIFT flow methods analyze the traffic sequences with large displacements. However, noise makes the large displacement optical flow incurred by the various techniques unsuitable for surveillance. The Tetrolet, a Haar-based wavelet transform, is proposed to denoise the large displacement optical flow. The simulation results demonstrate that the proposed Tetrolet transform is better suited for SIFT flow than Brox flow in terms of PSNR.

Download

Share

COinS