TRANSFER LEARNING BASED AUTOMATED CROP ANALYSIS

Abstract

In our project work, we have implemented the notion of transfer learning for classification of fruits, crops and vegetables. Every year the quality of large amount of crops and fruits deteriorates because of the lack of essential nutrients provided to them. In many areas, more than one crop are harvested in a single field which makes it very difficult for the farmer to provide proper nutrients to each and every type of crops as every crop requires specific kind of nutrients. To overcome this scenario, we have implemented the method of transfer learning using pretrained models to segregate the crops based on the features like size, colour, quantity and so on. Two pretrained deep learning models namely ResNet50 and MobileNet are fine-tuned appropriately to evaluate the quality of fruits and crops. For the evaluation of these fine-tuned models, we collected datasets of 15 different classes of fruits and crops. The dataset of each class consists of approximately 500 images. Based on these datasets, both the models were evaluated. The result of the models shows that ResNet50 model achieved the higher test accuracy as compared to the MobileNet model. The hardware implementation of our project includes the deployment and testing of our model on the Jetson Nano Developer kit in real time. Also, it includes calculating how long it takes the setup to compute a single frame at the time of real time testing.