발간년도 : [2021]
| 논문정보 |
|
| 논문명(한글) |
[Vol.16, No.5] Comparative Study on Performance Improvement Methods in CNN Based on Time Series Data |
|
| 논문투고자 |
Baek Ki Kim, Sangmin Suh |
|
| 논문내용 |
Human activity recognition (HAR) is a classification task based on time series sensor data, which have traditionally been performed with machine learning methods like a support vector machine (SVM). With development of deep learning, the task also is carried out using recurrent neural networks. Recently, it was confirmed that the time series data can be used for convolutional neural networks (CNN), where the time series data is transformed into an image beforehand to be used in CNN. Author has suggested several methods, i.e., edge enhancement, line width control, and color code optimization, for an accurate task in the time series based convolutional neural network. The edge enhancing method results from that recognizing things mean seeing the shape of the things. With the intuition additional edge information is applied to the image as an additional channel. Moreover, in the image of time series data, the optimal line width may be existed to achieve the best performance, which includes fast settle time and high accuracy. This motivates the line width control method. The color code optimization is motivated by a complementary color. The color code optimization can be obtained by maximizing the color distances in the color space. The complementary color makes images more clear and shape, which results in higher performance of the neural network. This paper compares and evaluates the three methods in detail. From the experimental results, in the image classification task of the time series data, the line width control is the most effective technique, and the color code optimization is also effective method. The line width control was the least effective because the raw information is time series data, i.e., line. However, if the raw data has spatial information such as images, the edge enhancement technique is much more effective to improve f1-score. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
