Enhanced Fashion-MNIST Classification Using a Hybrid VGG-16-DenseNet121 Architecture

This study aims to explore the effectiveness of a hybrid model combining the VGG16 and DenseNet121 architectures for image classification tasks on the Fashion MNIST dataset. This model is designed to leverage the advantages of both architectures to produce richer feature representations. In this study, the performance of the hybrid model is compared with several other architectures, including LeNet-5, VGG-16, ResNet-20, ResNet-50, EfficientNet-B0, and DenseNet-121, using various optimizers such as Adam, RMSProp, AdaDelta, AdaGrad and SGD. The test results indicate that the Adam and SGD optimizers deliver excellent results. The VGG16 + DenseNet121 hybrid model achieved perfect training accuracy 100%, the highest validation accuracy 94.65%, and excellent test accuracy 94.16%. Confusion matrix analysis confirms that this model is capable of correctly classifying the majority of images, although there is some confusion between classes with visual similarities. These findings affirm that a hybrid approach and the appropriate selection of optimizers can significantly enhance model performance in image classification tasks.