Multi-class image classification of fruits and vegetables using transfer learning techniques

Abstract

The need for an e cient fruit and vegetable identi cation and classi cation is always important and bene cial for not only the agricultural department or food processing industry, but also to the low level retail stores and supermarkets where fruits and vegetables are sold. Building an e cient automated tool is very much required. In order to build this application, an e cient and e ective classi cation model has to used, which can classify 1000's of fruits and vegetables in seconds. The purpose of this study is to nd the best classi er model which can be used to build this automated application. While many advancements have been made in recent years, many methods still struggle from prolonged training and testing time and even signi cantly more number of false positives after classi- cation. Thereby, in this paper, a review and experiments on 7 di erent available transfer learning models such as VGG16, ResNet50, MobileNet, DenseNet, InceptionV3, xception and InceptionResNet is conducted and compared by there accuracy, precision, F1 Score and training time, so that an e ective and e cient automated classifying system can be built in future. Along with this, a self-designed CNN model is trained and tested. The experiments are conducted using Fruit 360 dataset of 120 classes. Initial phase of this study involves training the models using subset of the dataset with 21 classes. VGG16 and ResNet50 are resulted as top 2 models. Thereby, the later phase of the experiment is conducted on these two models on the whole dataset with 120 classes. The overall results show VGG16 is the best model with 99% of training accuracy and 95% of testing accuracy.