Comparing Deep Convolutional Networks and Capsule Networks in Image Classification Tasks
Dr. T. Charan Singh, Suresh Bhukya
This research compares the performance of Deep Convolutional Networks (DCNs) and Capsule Networks
(CapsNets) in image classification tasks, evaluating their accuracy, training time, and inference time
across three benchmark datasets: MNIST, CIFAR-10, and ImageNet. DCNs, such as ResNet-50, VGG16, and AlexNet, have established themselves as robust solutions for image classification, excelling in
feature extraction and scalability. However, they often exhibit sensitivity to spatial variations due to
pooling layers. In contrast, CapsNets, with their novel capsule-based architecture and dynamic routing
algorithms, show promise in preserving spatial hierarchies and improving accuracy, particularly in
complex datasets. Despite achieving higher accuracy, CapsNets require significantly more computational
resources, with longer training and inference times compared to traditional DCNs. This study highlights
the trade-offs between the advanced spatial encoding capabilities of CapsNets and the efficiency of
established DCN architectures, offering insights into their practical applications and future research
directions.
