Design of EGTBoost Classifier for Automated External Skin Defect Detection in Mango Fruit

The presence of overlapping mangoes and leaves makes the segmentation process inaccurate. In addition, the variation in colour, texture, shape, temperature and shadow effects of the real conditioned image makes segmentation even more complex. The presence of outliers would further affect the classification accuracy. Many studies have reported the deep learning-based technique, but those methods failed to provide an effective solution due to the high processing stage and limited data usage. Hence, this work introduces a machine learning-based technique that reduces the problem of time-consuming predictions. This work aims to estimate the defected region of mangoes by enhanced segmentation and optimal feature selection to enhance classification accuracy. To detect mango fruit defects, initially, the collected mango images are pre-processed to smoothen and reduce image noise. This is achieved using the guided Gabor bilateral filter; the technique can reduce image noise without information loss. Then the obtained pre-processed image is segmented by considering the defect as a region of interest. The segmentation is achieved using the fuzzy level set method (FLSM), which creates clusters for an image’s dynamic variation. Then, the features are extracted using the dual-tree complex transform (DT-CT) and the optimal features are selected using the metaheuristic algorithm adaptive tunicate swarm optimizer (ATSO). The obtained optimal features are used for the detection process, which uses an Extreme Gradient Tree boost classifier (EGTBoost) classifier and the output is generated using vote-based classification. This classifier accurately classifies the diseased and healthy mangoes. The experimentation is carried out on the Kaggle and the real-time datasets. The accuracy and precision values achieved by the proposed model are 0.969 and 0.986 on the Kaggle dataset, respectively.