IDNet-A: Variant of DenseNet with Inception-Family
Cheol-jin Kim, Young-Guk Ha · 2020
A lot of interest in deep learning and advances in computer hardware (especially GPU) has recently led to many studies on network architecture in various fields. With these studies, the technology using machine learning in various fields shows good performance. Especially in the field of computer vision, solutions using convolutional neural networks (CNNs) are becoming overwhelming, with much better performance than solutions using image processing algorithms. In image recognition of computer vision, various network architectures using CNNs have been introduced and developed. In reference to previous studies, we introduce IDNet-A in this paper, which combines two impressive and powerful networks (DenseNet and Inceptionfamily). We studied how to increase the size of network to get good performance with increased representational power. We applied the Inception Module concept of Inception-family and Dense Connectivity of DenseNet to IDNet-A to efficiently increase the size of the network at a reasonable cost. As a result, we can train well in deep network architecture. We constructed several models with different hyperparameters and experimented with CIFAR datasets. Finally, IDNet-A, which we introduce in this paper, increases the size of the network by increasing the depth and width appropriately and achieves good performance with fewer parameters compared to other networks.