P‐217: Late‐News‐Poster: Convolutional Neural Network‐based Multi‐touch Detection Technique on Learning from Class‐imbalanced Dataset
Hyeon-Seok Yoon, Saehyun Ahn, Keon-Woo Kang, Il-ho Lee, Sang-jin Park, Yun-a Ma, Suk‐Ju Kang · SID Symposium Digest of Technical Papers · 2020
The touchscreen has a lack of accuracy in discriminating between multi‐touch and single‐touch signals when touch signals are too close to each other. This causes the machine's incorrect expectation of the intended touch coordinates. We propose a novel convolutional neural network (CNN)‐based multi‐touch detection to increase the accuracy for detecting the number of touching points. Additionally, the modified loss function is proposed to train the proposed neural network for the class‐imbalanced dataset. The novel loss function is optimized by considering the number of data according to the class, and then, the detection accuracy is increased compared with the conventional approach. Finally, we compute the coordinates of the touch signal by using the result of CNN. Experimental results show that the detection accuracy of the proposed network with our novel loss function was 93.67%.