<![CDATA[Image preprocessing is a critical stage in computer vision systems as it directly influences analysis and classification performance. Anisotropic Diffusion (AD) and Bilateral Filtering (BF) are widely used preprocessing techniques proven effective in reducing noise while preserving edge information, though each has limitations—AD may cause oversmoothing, and BF can be less effective against impulsive noise. This study employs three preprocessing approaches: (1) AD with a blending ratio of 70% original image and 30% diffused result, (2) BF with the same ratio, and (3) a hybrid AD+BF method combining both sequentially, where AD is applied before BF. At each stage, the processed image is proportionally blended (70%:30%) with the original to balance edge preservation and noise reduction. Experiments were conducted on plant disease classification using the MobileNetV2 architecture with the New Plant Diseases dataset containing 10 classes (560 training and 140 validation images per class). Results show that the hybrid AD+BF approach achieved the highest accuracy (99.21%) and F1-score (99.21%), outperforming standalone AD (98.86%) and BF (99.14%). The optimal parameters ( mm dan SUV) offer practical guidance for implementation. These findings provide empirical evidence supporting proportional blending as an effective preprocessing strategy for deep learning-based plant disease classification.]]>
Copyrights © 2025
