Abstract:
With the advancement of scientific technologies, manual identification for freshness of vegetables is not practical at all as it is time consuming and inaccurate. To solve this issue Implementing a machine vision-based system that can quickly discover the freshness of vegetables through images makes the accuracy level according to our aspiration. Dataset was collected by manually clicking pictures of our selected vegetables. To create a customized dataset, we carefully chose new conditions for three selected vegetables, ‘eggplant’, ‘okra’ and ‘cauliflower’, from the market and prepared the dataset. After collecting six thousand of basic images, we sorted them into 3 categories: fresh, aged, rotten which was decided by physical attributes and image capture time (Day). Association between raw dataset and augmentation, we tried to increase the amount of dataset where we will reach an effortless position to train a dataset. Transfer learning models that are used in the preliminary stage were InceptionV3, Xception, MobileNetV-2, DenseNet201. After practical implementation we got our desired result from DenseNet201. Though all these models give us decent accuracy level but after training DenseNet201 we were able to achieve the accuracy of 97.72%. This study will tremendously diminish the labor cost, and vegetables identification could also be made spontaneously.
