Abstract:

Aiming at the problem that the images obtained by the high-speed camera on the surface of the cylinder during the cotton fiber carding process cannot be recognized by the human eye, algorithms that combine image processing and deep learning were employed to assist human identification through a series of detection processes. The image data was derived from the high-speed video camera data of the carding process of the cylinder surface under the moving cover of the card. The specific implementation process was to first extract the denoising residuals from the image through a multi-level wavelet convolutional neural network, then use the deep convolutional networks for super-resolution reconstruction, and finally use a multi-scale edge detection and enhancement algorithm under strong noise to sketch the fibers. Through the processing of these three steps in the algorithm, a clear fiber image recognizable by the human eyes was obtained. Feature-enhanced image samples were used to train the cycle-consistent adversarial network to obtain more continuous and clear fiber extraction results. The results from the research demonstrate that the proposed processing procedure improves the detection and recognition effect of fibers during carding, and provides a new idea for the research in the field of carding.

Key words: cotton fiber carding, fiber image, fiber recognition, multi-level wavelet convolutional neural network, multi-scale edge detection