CUDA Accelerated and Optimized Bilateral Filtering for High-Resolution Food Image Enhancement and Quality Analysis

Abstract

High-resolution images of food tend to exhibit noise, uneven lighting, and varied textures that may impact the accuracy of food quality analysis performed automatically. The traditional bilateral filtering algorithm yields good results but suffers from high computational cost and thus is not recommended for use in real-time processing tasks. This paper suggests a framework for CUDA implementation of bilateral filtering to achieve high-quality enhancement and low computational delay. This work proposes to perform bilateral filtering on the GPU by means of parallel processing in which each CUDA thread would operate independently over image pixels. To increase computational efficiency, shared memory usage, coalesced memory accesses, asynchronous transfers between host and device memory, and kernel optimization techniques are used. The set of 500 high-resolution RGB food images with resolutions ranging from 2048×2048 to 4096×4096 pixels were tested under noise conditions. The performance measures used were noise reduction, edge preservation, texture preservation, PSNR, SSIM, precision, F1 score, computational time, FPS, and GPU speedup. CUDA-based BF produced 93.6% noise reduction, 91.8% edge preservation, 89.4% texture retention, and 91.6% performance in total. The algorithm provided 92.8% PSNR, 94.6% SSIM, 93.2% precision, and 93.0% F1 score. In terms of computational evaluation, it resulted in a processing time of 16 milliseconds per image, which equals 62.5 FPS and is 26 times faster than CPU BF. In scalability tests, stable GPU performance was shown, where the processing time grew from 11 milliseconds for 2048×2048 images to 22 milliseconds for 6144×6144 images. It can be seen that CUDA-based BF can increase the computational efficiency of high-resolution food-image enhancement significantly while keeping valuable structural and textural data. This method can provide tremendous benefit in food inspection, quality control, sorting, and grading applications. Future studies should be concentrated on researching the issues of adaptive filtering parameter values, illumination problems, multispectral imaging, and use of energy-efficient edge GPUs.