Identification of foreign fibers of seed cotton using hyper-spectral images based on minimum noise fraction

Abstract: In order to improve the recognition accuracy of seed cotton foreign fibers, the identification method in hyper-spectral images based on minimum noise fraction (MNF) was proposed and applied to feature extraction to reduce the dimension of multispectral images. This method can reduce the numbers of hyper-spectral data, and made the images noise reduce to the minimum and also reduce the computational requirements for subsequent processing. This paper selected white foreign fibers and cotton, which were in small discrimination, as the research object with 512 bands in the wavelength range of 400-1 000 nm. The spectral subset was selected according to the spectral curve, and then reducing dimension and denoising by using analysis method of MNF. The best component image was selected from the first four component images of MNF transformation by manual visual evaluation. The methods of image processing including median filtering, gray change method and so on were used to determine the best image segmentation and then extract the different fibers. Experimental results show that, for more than 5 kinds of different fibers, the recognition rate of the method reached up to 91.0%.