Abstract:
Objective. In this paper, a novel dataset has been collected in accordance with Pakistani
needs and is used to develop an architecture for the quality assessment of fruits and vegetables.
Approach. The dataset contains images captured under uncontrolled conditions with respect to
illumination, temperature, humidity, image resolution, image aspect ratio, angle of capturing
images and background. Images captured contain items individually as well as in groups. To the
best of the knowledge gathered, this is the first of its kind dataset. This dataset is then
preprocessed. Among usual preprocessing techniques, an aspect ratio adjustment algorithm has
been introduced. After preprocessing, the data is used to train multiple models (AlexNet, VGG-16,
ResNet-50, Fruits-360 Model and a proposed model with relatively lesser depth). This performs
recognition of fruits and vegetables and endorse the validity of the dataset. Going further, the
dataset is then prepared for quality assessment with three quality labels for each fruit/vegetable:
Eatable, Partially Rotten and Rotten. Quality assessment is then performed using pre-trained VGG-
16 through transfer learning, adding a fully connected network and fine-tuning the model. Main
Results. The highest recognition accuracy on the validation set is 98.9% and the highest validation
accuracy for quality assessment is 92.9%. Significance. Outcomes of this research demonstrate that
dataset collected under an uncontrolled environment can be used for recognition of
fruits/vegetables with remarkable accuracies. Moreover, quality assessment of fruits/vegetables is
performed accurately with the same dataset using deep learning and three quality labels.
