Data Augmentation Method for Improving Blurred Image Recognition Rate
Shiori Ishikawa, Miho Chiyonobu, Sayaka Iida, Masami Takata · 2023
This paper aims to improve the recognition rate of a recognizer for detecting cracks on concrete surfaces from image data. When developing a recognizer, it is common to use a clean image taken from the front as training data. Therefore, when blurred images that were not taken cleanly from the front are used as test data, the recognition rate decreases. To improve the recognition rate, the training data is mixed with processed images. Experimental results show that adding blurred images to the training data improves the recognition rate for blurred images, even when the images are blurred in different ways.