Detecting defective apples through advanced computer vision and Legendre Multi Wavelet Neural Networks

Large scale production and demand for apples of excellent quality to satisfy consumer and market standards have made an accurate, dependable and consistent grading system essential for the post-harvest process. Traditional manual visual grading of fruits poses substantial challenges to the agricultural business owing to its labour-intensive nature and the inherent diversity in inspection and categorization processes. In this study, Detecting Defective Apples using Advanced Computer Vision and Legendre Multi Wavelet Neural Networks (DDA-ACV-LMWNN) is proposed. Initially, the input image is collected from Internal Feeding Worm (IFW) database. The collected images are pre-processed using Multi observation Fusion Kalman Filter (MOFKF) to eliminate noise from the input image. Then, the pre-processed imageries are fed into the segmentation using Localized Sparse Incomplete Multi-View Clustering (LSIMC) method to extract the Region of Interest (ROI) and foreground. The segmented images are supplied to the Legendre Multi Wavelet Neural Networks (LMWNN), which detect and classifies the apples as Healthy and Defected. Generally, LMWNN does not expose any adoption optimization approaches for detecting optimum parameters to assure accurate defective apples classification. Hence, Tyrannosaurus Optimization Algorithm (TOA) is used for optimizing LMWNN which accurately classifies the defective apples. The proposed DDA-ACV-LMWNN is implemented in Python and the performance metrics like accuracy, precision, specificity, recall, F1-score, AUC are examined. The effectiveness of proposed DDA-ACV-LMWNN approach attains 16.60 %, 24.01 % and 26.60 %, higher accuracy; 20.32 %, 28.50 % and 32.21 % higher precision when compared with existing techniques like Predicting apple plant diseases in orchards using machine learning and deep learning approaches (PAPD-ML-DL), Detection of bruises on red apples using deep learning techniques (DBRA-DL) and Hyper spectral imaging coupled with deep learning technique for visualization and identification of early bruises on apples (HIC-DL-EBA) methods respectively. • Proposed DDA-ACV-LMWNN for accurate detection of defective apples. • Used MOFKF to pre-process and denoise apple images from the IFW database. • Applied LSIMC for effective segmentation and ROI extraction. • Optimized LMWNN with TOA to enhance defect classification performance. • Achieved higher accuracy and precision than PAPD-ML-DL, DBRA-DL, and HIC-DL-EBA.