This project is intended to learn Deep Learning core concepts. I've also deployed the same project by using CNN algorithm and tensorflow API. You can find the deployed project here.
In this project, I've implemented a Food/Non-food binary classification using Artificial Neural Network (ANN) from scratch. This project is taken from Kaggle Datasets but the dataset is also available at this link. The images used for the dataset are scaled down to size of 64 X 64 X 3 and the model has 3 hidden layers and one output layer using sigmoid activation function. After evaluating the model gave an accuracy of 75.0%.
The dataset used for the project is Food-5K. This is a dataset containing 2500 food and 2500 non-food images. The whole dataset is divided in three parts: training, validation and evaluation. The naming convention is as follows:
{ClassID}_{ImageID}.jpg
ClassID: 0 or 1; 0 means non-food and 1 means food.
ImageID: ID of the image within the class.
The first step in data preprocessing is to convert the image into array of numbers which was done using OpenCV's cv2.imread() function. After the conversion, the image was then resized to shape of 64 X 64 X 3. Then the array was flattened to achieve a shape of 64*64*3 = (12288,m) where "m" is the number of training examples. The last step in data preprocessing is to normalize the values in the flattened array. The normalization is done by dividing the array by 255.
While initializing the parameters, I've used "He weight initialization" which simply multiplies the weights by sqrt(2 / (No._of_neurons_in_previous_layer)).