ADAPTIVE FILTER FUSION AND OPTIMIZED HARDWARE DESIGN FOR IMAGE DENOISING

Abstract

Image denoising is vital for retaining visual information for tasks such as object detection, classification, and image enhancement. The practical issue with image denoising is one of balancing noise reduction with maintaining key image features, such as edges and textures. The contributions of this thesis are outlined in twin aspects, being a software-based adaptive filter fusion solution of color image denoising, and an optimized hardware implementation of an adaptive filter that performs real-time denoising of gray-scale images. More specifically, this research focused on algorithmic adaptability and hardware efficiency in various noise types such as Gaussian and salt-and-pepper noise. The first part of the thesis presents an adaptive fusion filtering framework for color image denoising under Gaussian noise for standard deviations of σ = 10 through σ = 50. The proposed framework employs a dynamic per-pixel fusion of three carefully selected adaptive filters: Least Mean Squares (LMS), Llncosh, and VSLMS Ang’s. These filters were chosen as they complement each other based on their strengths in texture sensitivity, edge preservation, and stability following benchmarking studies where we analyzed the performance. The merger utilizes an inverse squared error−based weighting scheme that spatially adapts as a function of the image pixel locations under denoising. The use of bilateral preprocessing enables improved edge-aware smoothing characteristics while retaining fine-level structural details. The adaptive filter fusion strategy improves both PSNR and SSIM image quality metrics significantly over using the individual filters indicated above, as shown through the testing on the CBSD68 dataset. Bias-variance analysis and run time profiling support the purpose and practicality of the proposed scheme. The second component examines the hardware modeling of a fixed-weight Least Mean Fourth (LMF) filter where denoising of grayscale images occurs in an impulsive noise environment. The Least Mean Fourth (LMF) filter has been selected as the approach is insensitive to outliers via the approach to adjusting the fourth power of the error. The LMF filter is trained off-line utilising grayscale images from the BSDS500 dataset, which features artificial added salt-and-pepper noise to emulate images with impulsive corruption. The adaptation employs fixed point arithmetic to provide for field programmable gate array (FPGA) deployment. A standard 3×3 convolution is employed with a streaming architecture to avoid buffering the entire frame and allows for line buffer-based processing. Wallace tree multipliers were used to minimize log delay in the datapath and allow synthesis timing closure without altering the output, and to accelerate the overall implementation. The final architecture achieves a greater than 90% reduction in resource utilization compared to the non- optimized multiplier-based design, yet with nearly identical output fidelity to the software reference.