Automated Fruits Quality Assessment through Computer Vision and Deep Learning

Abstract

In agricultural field, automation is essential to raising a country's standard of living, economic expansion, and productivity. A large part of a person's diet is made up of fruits, and finding good fruit in the market can be challenging at the moment. Sorting and grading them has a big impact on customer preferences, choices, and market value. This affects both home consumption and export markets. While human grading and sorting is feasible, it is ineffective, arbitrary, and unreliable, increasing costs and the likelihood of errors brought on by external factors. Rather than employing embedded systems (sensors) or the labor-intensive manual methods of grading each fruit and vegetable by hand, which would take longer to complete, we chose to use a highperformance Android application for faster deployment in order to expedite data identification and improve usability. This comprehensive study investigates the application of deep convolutional neural networks (CNNs) in automating fruit quality assessments using computer vision techniques. With a dataset comprising 14,400 original fruit images, the research employs advanced augmentation methods to enhance dataset diversity, generating multiple images. The study evaluates the performance of five prominent CNN models ResNet50v2, VGG19, EfficientNetB0, InceptionV3 and a hybrid model of DenseNet121 & EfficieintB6. The hybrid model has achieved the highest accuracy of 92.43% but in terms of efficiency the hybrid model required much more computational power and as well more inference time which make it less efficient for mobile devices. On the other hand, EfficientB0 model which is far more superior in terms of efficiency was achieved 92.14% accuracy which is very close to what the hybrid model has achieved. While MobilenetV2 achieved slightly lower accuracy of 91.64% and it is as well very efficient as EfficientNetB0. InceptionV3 was also highly accurate, with accuracy of 90.63%. The accuracy of ResNet50v2 and VGG19 were slightly lower, measuring at 89.19% and 89.62%. Based on these EfficientNetB0 is the overall best model in terms of both accuracy and efficiency.