RGB Channel Combinations Method for Feature Extraction in Image Analysis

Abstract

Latest image analysis deep learning algorithms use diverse methods to extract features from images based on the Convolution Neural Network (CNN). CNN has a convolution layer consisting of RGB as three overlapping channels in the feature extraction process, and such architecture enables the backbone network to flow without losing each hue information. Therefore, 3D input data consisting of 3 channels to process the RGB channel consists of a large-scale neural network with many layer blocks. This processing method exhibits high accuracy. However, in terms of practicality, it results in big inefficiencies such as memory overhead and computational overhead. This study proposes the RGB Channel Combinations Method for Feature Extraction in Image Analysis to resolve such inefficiencies. This is a method that converts the RGB value into one tensor structure by combining each weight and bias and makes it possible to pass through the backbone network without damaging hue information. Based on the experiment results, it is confirmed that the accuracy decreased by 1.42% compared to the pre-existing method, but the number of parameters used by the input layer decreased. It is confirmed that the pre-processing used in the proposed method gained an additional computational overhead, but the number of input parameters decreased to 1/3 in the feature extraction stage performed afterward. As the proposed method applies to all image analysis algorithms, its expandability is extremely high and can process a large amount of image data.