Computational comparison of image preprocessing techniques for plant diseases detection

There is an abundance of deep neural network models for plant disease detection. Prior to applying these models, image preprocessing techniques are applied in order to improve the detection results. However, there is a lack of computational comparisons on the application of different image preprocessing techniques before applying object detection algorithms for plant disease detection. This paper aims to fill this gap by presenting a computational comparison of seven different image preprocessing techniques (auto-orientation, object isolation, resizing, grayscale conversion, static crop, contrast adjustment, tiling) applied prior to the execution of two state-of-the-art object detection algorithms, one single-stage detector, YOLOV5, and one two-stage detector, Faster-RCNN. We investigate whether or not these preprocessing techniques improve the accuracy, training time, and inference time, of plant disease detection. Apart from comparing these techniques solely, we also perform combinations of the preprocessing techniques. The PlantDoc dataset was used for this experimental study. Computational results show that the best method improves the mean average precision by 9% and 3% for YOLOv5 and Faster-RCNN, respectively. Finally, the combination of all seven preprocessing techniques yields an improvement of about 13% in the mean average precision of both object detectors.