Studying the Performance of Transfer Learning on CNN Models for Fruit Sorting

The field of computer vision has made some significant breakthroughs in the past years one of which has led to convolutional neural networks (CNN) being state-of-the-art algorithms for computer vision tasks. This breakthrough led to the development of different types of CNN architectures by researchers. These CNNs are typically evaluated on the popular ImageNet competition thus making it the benchmark for performance analysis of CNNs. This research leverage some of the popular pre-trained models; MobilenetV3-small, Resnet50, and VGG-16 for the classification(sorting) of fruit images according to their appearance. The models are fine tuned to obtain the best performing model to solve the fruit sorting problem which consumes a lot of resources in the form of labor, time, and finance. The results show that transfer learning can be successfully applied to sorting fruit and the model with outstanding performance is VGG16, followed by ResNe $t$ 50, and lastly MobilenetV3-small with an average accuracy of 95%, 76%, and 63% respectively. This research shows a comparative analysis of VGG-16, MobilenetV3-small, and Resnet50 of their performance during sorting. In addition, this research also investigates the performance of the model when the number of fruit classes is increased.