ConvNext: A Contemporary Architecture for Convolutional Neural Networks for Image Classification
Agastya Todi, Navya Narula, Moolch Sharma, Umesh Gupta
Abstract
Convolutional neural networks (CNNs) have established themselves as the industry standard for image classification jobs. Convolutional neural network architectures have been created and enhanced recently to improve accuracy and decrease computational complexity. Convolutional Neural Networks' effectiveness declined as transformers became more widely used in computer vision. The transformers, such as vision transformers, outperformed the antecedent CNNs in image classification tasks requiring computer vision. However, in 2022, Facebook researchers created a collection of convolutional neural networks that performed better than vision transformers in the Image Classification experiment. A ResNet was used to build this network, subsequently “modernized” to make it more resemble a Vision Transformer. Following this, the curated model was trained and evaluated on the ImageNet1k and ImageNet22k datasets for Image Classification, yielding several novel results. ConvNext was the name of this network, and it has many variations. On the CIFAR-10 dataset, we assess the effectiveness of the ConvNext models in this work. We demonstrate the advantage of the ConvNext model in terms of accuracy by comparing the results with those from other cutting-edge models.