Hyperparameter Optimization of an Image Classification Model for Beef and Pork Using Convolutional Neural Networks

Abstract

Deep learning classification network in one case, has different classification capabilities than the network in another. The classification method of deep learning using CNN has specific hyperparameters that can be adjusted to have good performance. These hyperparameters include the number of convolutional layers, the number of neurons in the convolutional and fully connected layers, kernel size, and activation functions. Deep Learning uses experimental principles in finding the best hyperparameter in various cases. The model architecture can be determined by choosing a different design. This research uses pork and beef images as the data for classification using CNN. The abstract textures of beef and pork may make it difficult for the CNN classification model to distinguish between them. Hence, 32 combinations of five hyperparameters were compared. It was found that these hyperparameters affect the model's performance. The best model has obtained 98,7% accuracy that uses 20 neurons both layers of the convolution was, kernel size of 5 × 5, ReLU activation function, and two fully connected layers with dropout 0.7 as a method of overfitting prevention. A significant difference also occurs in the application of the activation function, in which ReLU has a better performance than tanh function to increase the model's prediction.