Convolutional Methods in Image Enhancement

Abstract

Convolution and deconvolution are essential in image processing for enhancement, analysis, and feature extraction. Convolution is widely used for filtering and edge detection, while deconvolution restores blurred images by recovering hidden details. These techniques are particularly important in space research, where high spatial resolution is coupled with 12–16-bit ADC, and optical quality is often degraded by stray light, making data interpretation challenging. Additionally, high-resolution images with large spatial dimensions and a high number of pixels, commonly encountered in space research, require significant computational resources, leading to slow processing times.
Our research focuses on optimizing convolution and deconvolution techniques using CUDA technology to accelerate processing. We developed a custom CUDA-based stray light removal method, achieving performance comparable to a previous C++ implementation while significantly reducing processing time through parallelization, resulting in an approximately 83% reduction in execution time and a runtime below half a second.
For deconvolution, we implemented multiple algorithms in MATLAB and CUDA environments, including Wiener filtering, Richardson–Lucy deconvolution, relevant regularization methods, and blind deconvolution. The Richardson–Lucy method, due to its iterative nature, is computationally intensive, which motivated its CUDA implementation. Leveraging GPU parallelization, we achieved substantial speed improvements – specifically, more than a 52% reduction in execution time – while maintaining result quality.
This paper proposes multiple deconvolution solutions for various image processing tasks and demonstrates the effectiveness and applicability of parallel programming in image enhancement algorithms. These contributions are particularly valuable for large-scale images and real-time applications.