| Over a long period, people have harvested preharvest cottons by manual sense mostly in China, which necessarily sways cottons grades due to subjective factors. To quicken steps for the purchase of preharvest cottons mechanically, the paper presented a thought of classifiers design for field sample preharvest cottons with bracteoles based on machine vision and multiple regress. Preharvest four-petal cotton images afield on dark background were acquired with seven grades and great efforts were made on the following pivotal technologies:1. Images segmentation. Image enhancement procedures such as median filter, pixel-to-pixel gray-scale transformation and morphological operation are generally used to solve some problems such as poor contrast or noise caused by inadequate and non-uniform illumination. In experiments, firstly, cotton templates with bracteoles were extracted from original images on dark background of noises. Approaches are as-follows: original gray images were transformed into binary images using rain-threshold, and cotton center binary templates were extracted. The enhanced cotton images based on top-hat operations were transformed into binary images using Otsu-threshold, and cotton border binary templates were extracted. And then, both templates were added. Secondly, cotton templates were extracted from cotton gray images with bracteoles and background of 0. Approaches are as-follows: the enhanced cotton gray images based on bottom-hat operations were transformed into binary image using Otsu-threshold, and cotton and its bracteoles were separated. Then morphology-based opening operation were performed using a big structure element to binary images, and the number of objects left in binary image were used to judge the connectivity of cotton petals. For low quality cotton with multi-connectivity, some cotton petal templates were extracted in turn from non-opening binary images. For high quality cotton with single-connectivity, max cotton petal templates were extracted from binary templates which original gray images were transformed into using Otsu-threshold.2. Features extraction. Statistical procedures such as mean and standard deviation were used to extract features such as size, color. According to Chinese literal government standards in the purchase of preharvest cottons, serial parameters based on machine vision were extracted as-follows: cotton templates areas with/without bracteoles, cotton yellow area of (R+G)/2-B with bracteoles, cotton intension mean and standard deviation with/without bracteoles. 3. Classifier design. Independent variables were as follows: let xl be the ratio of cotton area without bracteoles to bracteoles area, let x2 be the ratio of cotton yellow area to cotton area with bracteoles, let x3/x5 be cotton intensity mean with/without bracteoles, let x4/x6 be cotton intensity standard deviation with/without bracteoles. In addition, let y be cotton grades as dependent variable. In experiments, total samples were separated into training data and test data, and two-element regress model based on training samples was established as follows: y= 3. 859 - 0. 254 x1+11. 595·x2, and R~2 = 0.755, F= 85.571, p=0. All suppositions of multiple regress model in theory were proved, the model is actual and reliable. When the left test samples were substituted into the model, classification precision of the model is 75%. Contributions of x1 and x2 are 50.4% and 49.6%. Hence, machine vision system can help standardizing and quantifying the inspection process of field preharvest cottons by promoting grading consistency and objectivity. |
