Özet:

Convolutional neural networks (CNN), a feedforward neural network model, are frequently used in image classification problems. New architectures have been developed by making some additions and changes in order to improve the performance on the CNN model, which generally consists of convolution layer, pooling layer and fully connected layer. By adding different numbers of hidden layers with sigmoid, tanh and ReLu activation functions to the CNN-based VGG16 and ResNet50 architectures developed in this study, images were classified by deep transfer learning technique and the performances were compared. The dataset used for classification is a multi-class dataset related to fruits, and a 10-neuron softmax classifier was used in the last layer of the established models. Entering the epoch as 10, results were obtained for four different metrics: accuracy, precision, recall, and f1-score. When the results are compared, it has been observed that the model created by adding a total of three layers, including two hidden layers with 256 and 128 neurons, and a softmax classifier layer with 10 neurons, to the ResNet50 architecture with sigmoid activation function, gives the best result with a classification accuracy value of %97.5. Finally, the results obtained for the four metrics were subjected to the Friedman and Nemenyi post-hoc tests, a statistical analysis was made and the relationship between the models was tested. It was concluded that the models created as a result of the test were related to each other.

Anahtar Kelimeler: