A lot of interest in A.I.(Artificial Intelligence) and the advance in computer hardware (especially GPUs), many studies on A.I Network structures are being conducted in various fields. Technologies using A.I. in various fields show good performance. E...
A lot of interest in A.I.(Artificial Intelligence) and the advance in computer hardware (especially GPUs), many studies on A.I Network structures are being conducted in various fields. Technologies using A.I. in various fields show good performance. Especially in the field of computer vision, solutions using convolutional neural networks (CNNs) have gained much better performance than solutions using image processing algorithms. And the solutions have become overwhelming. Various network structures have been introduced through competitions and studies in the field of computer vision, and among them, various studies have been conducted on a network structure based on a CNNs with good performance. We have studied various network structures based on CNNs introduced in the previous study. And We studied how to increase the network size (depth and width) efficiently to obtain good performance. Increasing the network depth and width in the network structure can increase the network representational power that the network can represent, which can improve performance. However, increasing the size without careful consideration causes an exponential increase in the number of parameters, which can lead to computational and overfitting problems. In addition, in a deep network structure, gradient vanishing/exploding or degradation problems that adversely affect learning may occur. These problems have a critical negative impact on network performance, making it impossible to increase the size of the network.
In this paper, we introduce a method of combining the Inception Module in densely connected network structure to solve the above problem. Since densely connected network structure is connected from one layer to all other layers, information from the front of the network is well transferred to the back. This makes it possible to construct the network structure deeply by alleviating gradient vanishing/exploding or degradation. This densely connected network structure has one drawback. Since it is a densely connected structure, when the width of the network is increased, the number of parameters increases a lot, which may cause a problem of computation and overfitting. So, we combined the inception module in densely connected network structure to efficiently increase the network width. The inception module extracts various features of an input value through filters of various scales, which can be used as useful information during learning. And it is possible to reduce the number of parameters and the amount of computation by adjusting the number of channels through the 1x1 Conv inside the module. By combining such an inception module with densely connected network structure, it is possible to increase the network size efficiently without significantly increasing the number of parameters and computation. It also extracts the various features required for learning, so that the parameters efficiency in densely connected structure can be improved, and good performance can be achieved with a small number of parameters. We introduce a network structure called IDNet-A that applies the idea we introduced in this paper. IDNet-A is a network structure that combines the inception module with DenseNet, densely connected network. And several IDNet-A models with different hyperparameters were constructed and tested with the CIFAR dataset to confirm the performance according to the network size. IDNet-A proposed by us increases the network size by appropriately increasing the depth and width of the network. And IDNet-A shows better performance than DenseNet.