Abstract : Food losses due to plant disease and pests have been an issue for many years. The early non-invasive detection and prediction of leaf diseases with modern technology are crucial in the contemporary world to sustain food security. Binarization is the primary technique for extracting significant information from an image. However, it could not always extract substantial information due to variations in the distribution of the image pixel histogram because of the delivery of light during image acquisition. Image acquisition from different environments and laboratory settings could potentially result in a highly non-uniform histogram distribution of leaf images. Extraction of Region of Interest from leaf images with non-uniform illumination has been challenging for the conventional thresholding and edge detection techniques. It is because the optimal threshold value for the highly varied histogram distribution is difficult to calculate dynamically. This paper proposes an Enhanced ELA incorporated with the Otsu method to improve the binarization process of the non-uniform histogram intensity distribution of the leaf images. The combination of Enhanced ELA with Otsu acquired the best performance in three measurement metrics: validation accuracy, convergence rate, and least overfitting compared to the benchmark models. CNN architecture is used for the classification of plant leaf images.