The paper has been published in IEEE UPCON22 and can be found here.
This repository aims to solve the task of fruit quality classification given very limited data. The dataset contains 2690 images of lemons which is too less to train a deep learning model, let alone use the model practically.
Fruit quality classification of defective fruits is an important task in the fruit processing industry. An automated system that could determine fruit quality by classifying it on the basis of surface defects is needed for enhancing fruit production, reducing the dependency on manual labour and reducing losses due to fruit quality deterioration.
We solve this task by first augmenting the data trying many variations of GANs - cGAN, WGAN and WGAN-GP with WGAN-GP generating the best images and giving the best results upon combining with the classification pipeline. The classification using transfer learning by trying different pretrained CNN models - VGG16, Xception, DenseNet, EfficientNet and ResNet with ResNet giving the best results of 92.11% testing accuracy.