Enhancing Image Classification Performance Using Multi CNN Feature Fusion Method

Abstract

This research aims to overcome general challenges in the field of image pattern recognition using a convolutional neural network (CNN), which is still faced with the complexity and limitations of image data. Achieving high accuracy is essential because it significantly influences the effectiveness and success of numerous areas. Although deep learning technology, especially CNNs, offers the potential to improve accuracy, it is still limited to the 70–80% range for achieving the expected level of accuracy. In this research, a fusion method was developed that combines pre-trained models using concatenation techniques to increase accuracy. By utilizing pre-trained models such as ResNet50, VGG16, and MobileNet-v2, which were then adapted to various datasets and cross-validation techniques, researchers managed to achieve significant improvements in accuracy. The results of this study show an improvement in the accuracy of the Fusion Multi-CNN model for various datasets. On the fashion dataset, MNIST managed to achieve an accuracy of 0.87840, while on CIFAR-10 and Oxford-102, the accuracy was 0.81260 and 0.84004, respectively.