Fruit Disease Classification and Localization Using Region Based Regression

Abstract

Fruit disease classification using computer vision techniques is completely viral upon machine learning capabilities. The complete analysis is based upon regression using region proposal networks and their optimization using Gradient Clipping methodology. Fruit diseases such as flyspeck, blotch, scab and rot are detected and classified using the Region proposal network regression. The recent works carried out on the fruit disease detection and classification on agricultural crops were gathered and surveyed in this paper. Finally the very recent work carried out using regression and computer vision techniques were identified and applied to the data collected here with 6231 images and 24 classes. The diseased and disinfected were filtered for training purpose into 3110 images as diseases and the rest as disinfected. The hyperparameters tuning optimization was able to fit only the random data images, instead gradient clipping resulted in the proper limit cropping of the diseased portions of the crop footage. To improve the training data stability regression was employed with this optimization to show the results obtained from 95.3% to 97.8%. Here, the RCNN classification using neural networks resulted in the overall accuracy of the fruit disease classification model to 95.83%, where the Gradient Clipping optimization resulted in the improvement of accuracy of model to 97.8%.