ANN-SVM-IP: An Innovative Method for Rapidly and Efficiently Detecting and Classifying of External Defects of Apple Fruits

Abstract

One of the highest concerns of worldwide fruit and vegetable enterprises is the rapid and accurate separation of fresh apples from damaged ones, where efficiency and cost-effectiveness are critical. Most recent studies have neglected the speed and cost of separation, both of which are crucial in the fruit and vegetable business, in favor of accuracy alone. Such an approach has a negative impact on these procedures. This paper attempts to address these issues by offering a novel method (ANN-SVM-IP) that integrates image processing (IP)-based segmentation and ML algorithms (ANN-SVM) for extracting and classifying exterior defects in apple fruits. The proposed strategy combines accuracy, rapidity, and affordable implementation cost. The first phase attempts to detect exterior defects in apples by applying two proposed convolution kernels that were capable of identifying damaged sections of apples. The second phase is designed to accurately and effectively classify five Apple fruit defects (Healthy, Full-Damage, Bloch, Rot, and Scab) an optimized ML (Machine Learning) algorithm, which relied combining ANN (Artificial neural network) and SVM (Support Vector Machine) techniques. This paper aims to provide a practical and systematic solution to an important problem in the field of agricultural automation. Therefore, as the results demonstrate, the integration of traditional machine learning techniques (artificial neural networks and support vector machines with image preprocessing) provides a cost-effective and efficient approach for detecting and classifying external defects in apples. The proposed model focuses on speed and scalability by relying on the combination of traditional machine learning techniques, which are characterized by low cost, high performance, and simplicity of implementation, which was confirmed by the results obtained (99.4% for separating different defects, 97% for defect extraction, and a speed of 3.29 milliseconds, sufficient to complete defect identification and classification for each frame). To ensure the reliability of the results obtained, all the techniques used in the trading were evaluated using several different evaluation mechanisms, including accuracy, precision, recall, F1-score, confusion matrix, and runtime analysis. The results were also compared with those of several of the latest techniques published in recent years. In addition to relying on a large database of more than 21,000 diverse images, it makes the proposed approach promising and feasible for practical applications.