Fruit classifier

This project is a simple fruit classifier trained with the Fruits 360 dataset, available on Kaggle.

Fruits 360 dataset

The Fruits 360 dataset contains 131 classes for a total of 82110 images. It was created from extracting frames of 20-second recordings of fruits rotating on a motor. All the images were rescaled to 100x100 pixels and the background of all images has been removed. Some image examples are:

Model

The CNN model is implemented in Tensorflow with Keras. The backbone used for transfer learning is the ResNet50, pre-trained on ImageNet. The top and output layers to be trained are:

• A fully connected layer with 1024 neurons, activaded by ReLU (Rectified Linear Unit);
• A fully connected output layer with 131 neurons, activaded by Softmax. This is used to get the probability of each one of the 131 classes.

Results

The metrics obtained with the test set are the following:

• Accuracy: 0.9017
• Loss: 1.3492