Image Segmentation Technique of Field Cotton Based on Color Threshold
WANG Ling,WANG Ping,CHEN Bing-Lin,LIU Shan-Jun,JI Chang-Ying
DOI: https://doi.org/10.3724/SP.J.1006.2010.00502
2010-01-01
ACTA AGRONOMICA SINICA
Abstract:The goal of cotton production in China is to improve corresponding rate of cotton quality grade, foreign fibers, adulteration, and cotton baling inconsistent phenomenon to decrease continuously. With the background, machine vision and pattern recognition technologies are introduced into traditional picking task to discriminate maturity degree and grade of quality of field cotton, which will solve the problem of picking cotton by the way from source, so that various cotton varieties can be adapted, pollution caused by agriculture chemicals can be avoided, labor cost can be reduced and agriculture cost can be decreased. In order to segment field cotton images exactly, we regarded cotton and its background as two classes and segmented them based on their color threshold. A total of 20 000 white, yellow, and stain cotton pixels and 20 000 background pixels of soil and cotton plant, including cotton bracteole, leaf, and branch, were extracted from typical under-ripe cotton images and ripe/over-ripe cotton images with various quality grades from 1 to 7. Color threshold of two classes of cotton and its background pixels were obtained in RGB, HSI, La*b*, and Hunter color space respectively; on the basis of which cotton regions were segmented from images; and HSI and La*b* color spaces were selected respectively by using S below 28, I over 108, L over 118, a* from 123 to 134, b* below 136 with less segmentation noise which would be removed based on morphological filter. The experiment results showed that 907 cotton images were segmented with an accuracy of 87.21% and 86.33% in HSI and La*b* color space respectively. The front images were segmented with an accuracy of 90.83% and 89.98% and the side images with an accuracy of 83.33% and 82.42%. Ripe cotton images were segmented perfectly in HSI color space while under-ripe cotton images in La*b* color space, and the speed-based segmentation method with threshold covering a wide area was preferable for field cotton surroundings.