Özet:

Abstract The automation of fruit classification represents a compelling application of computer vision. Historically, fruit classification methods have frequently depended on manual procedures that rely on human visual perception. However, these methods are characterized by their tedious nature, lengthy time requirements, and lack of consistency.The primary criterion for classifying fruits is their external shape and appearance.In recent years, the fruit industry has witnessed a growing utilization of computer machine vision and image processing techniques. These technologies have proven to be particularly valuable for various applications within the industry, such as quality inspection and the sorting of fruits based on colour, size, and shape. Research conducted in this field suggest that the utilisation of machine vision systems holds promise in enhancing product quality and alleviating the need for manual fruit sorting processes.This paper presents the design and development of an intelligent automation system for the grading of mango fruit.The fruit segmentation process is initially conducted using the U-NET model, which incorporates feature extraction from EffecientNet+ResNet101. The utilisation of the Artificial Bee Colony Optimisation algorithm has been employed in order tominimise and optimise the feature vector.The categorization of fruit quality has been accomplished by considering surface defects and maturity classification. Therefore, in order to classify defects, the 1D-CNN has been trained using optimal segmented features of abnormalities.  To enhance the performance, the proposed ACO_1DCNN model is employed to optimise the classification accuracy. The proposed classifier is utilised to categorise the maturity levels as ripe, partially ripe, and unripe.Ultimately, the defect and maturity output have been employed as determinants for categorising the quality as either good, average, or bad. The experimental results demonstrate that the ACO_1DCNN model attains an accuracy of 97.95%, precision of 95.97%, recall of 95.59%, F1-score of 95.76%, AUC score of 99.42%, and MCC of 94.34% when evaluated on the testing dataset.During the training process, the model attains an accuracy of 97.62%, precision of 95.83%, recall of 94.96%, F1-score of 95.33%, AUC score of 99.56%, and MCC of 93.69%.

Anahtar Kelimeler: