잡음 패턴 인식을 위한 양방향 연산 메모리 기반 퍼지 지도 학습 방법

Abstract

Recently, deep learning techniques have been applied to image analysis and recognition fields with various features. In particular, CNN (Convolution Neural Networks) is a deep neural network that has excellent performance in image processing and recognition performance among various deep learning models. However, in order to have excellent performance in recognition of patterns or images with a lot of noise or severe damage, many learning patterns are required and the labeling work for the learning patterns must be accurate. In general, it is efficient to apply CNN, a deep learning technique, to images with quantified features. However, when applying CNN to images with noise or severe damage, a lot of learning data is required. Therefore, it is efficient to select a supervised learning algorithm based on data analysis.

In general, supervised learning methods in deep neural networks or deep learning require many training data pairs to improve the recognition performance of patterns. However, when learning patterns that are noisy or severely damaged, the classification performance of learning deteriorates or learning does not occur. Therefore, in this paper, we propose a fuzzy supervised learning method based on bidirectional associative memory that can improve the classification performance of learning and enhance the recognition performance with a small number of training pattern pairs even when there is noise or some features are damaged.

In the proposed learning method, the fuzzy supervised learning structure based on the bidirectional associative memory is a complete research structure. In the proposed learning structure, the learning of the input layer and the associative memory layer is performed using a recurrent learning structure. And supervised learning is applied between the associative memory layer and the output layer to adjust the connection weights and bias terms.

And in the input layer and associative memory layer, BAM is applied to restore damaged features or remove noise from noisy features to form quantified data as input to the output layer. And the quantified data is transferred to the output layer to apply fuzzy map learning.

In this paper, we implemented and experimented with Visual Studio C# on a PC equipped with an Intel(R) Core(TM) i7-8400 CPU and 16GB RAM. In this paper, we compared and analyzed the recognition performance of the proposed learning method with the existing fuzzy supervised learning method, error backpropagation algorithm, and CNN among deep learning techniques.

In order to analyze the recognition performance of the proposed method, this paper conducted recognition experiments on patterns with noise. Noise is generally referred to as noise. Representative methods for experimenting with noise include salt-and-pepper noise and Gaussian noise.

When the features of numeric or English letter patterns are normalized but there is a lot of noise in the patterns, the image processing process is performed to remove the noise and then the learning pattern is constructed. However, in the process of removing the noise, the features are lost or the learning patterns containing noise are constructed, so when applied to machine learning or deep learning techniques, the classification performance is sometimes degraded. Therefore, in this paper, we propose a fuzzy supervised learning algorithm based on homogeneous associative memory that can effectively classify and recognize features with a small amount of learning data even when noise is included. In order to analyze the performance of the proposed learning method, the recognition performance was compared and analyzed with the existing fuzzy supervised learning method, error backpropagation algorithm, and CNN, and it was confirmed that the proposed method improved the recognition performance with a small amount of learning data for various noisy patterns. 14