| The development goals of cotton production were including, such as the corresponding rate of cotton quality grade improved greatly, the foreign fiber problem, the adulteration event, and the inconsistent phenomenon of cotton baling decreased continuously, etc. The main measures of solving these problems were supervising strictly picking, purchase, and process cotton and decreasing foreign fiber content. Modern technology, for example, machine vision and pattern recognition, was applied in traditional picking task to discriminate ripeness and quality grade of raw cotton, and thus laying a foundation for picking robot to exert exactly. Using robot to pick cotton could adapt various cotton varieties, avoid pollution caused by agriculture chemicals, and reduce labor waste and agriculture cost. The main and creative achievements as follows:1. Raw cotton and its background were regarded as two classes and segmented based on competitive learning network. The training data with class label was classified into two classes based on K-means clustering in HSI, Lab, Ohta, RGB color space, and the error rate of the training data was lowest in RGB color space-particularly B value. The competitive learning network with B of the training data was trained onetime in RGB color space.907cotton images were segmented with an accuracy of92.94%based on the competitive learning network and morphological filtering. With the help of prior knowledge and supervisory learning, the image segmentation arithmetic obtained high precision and efficiency without iterative and over-fitting of K-means clustering as well as high computing cost.2. The spectrum feature set used to describe boundary of raw cotton was extracted in frequency domain. A sequence of coordinates of boundary contour of cotton were obtained by using8-neighbour tracking, from which64spectrum features without0.9correlation coefficient were extracted based on Fourier transform. Single factor anova analysis showed that the differences of the average of most spectrum features of ripe and unripe cottons were significant at the0.05level.The geometric structure feature set used to express cotton shape was extracted in spatial domain. Cotton region was circumscribed or cut by a set of concentric circles, from which15geometric structure features with more0.9correlation coefficient were extracted based on ratio method, including global feature, three radial cut region features, five circumferential cut region features, two radial cut line features, three circumferential cut line features, and box-counting dimension, Single factor anova analysis showed that the differences of the average of most geometric structure features of ripe and unripe cottons were significant at the0.05level.Ripeness and under-ripeness of raw cotton were regarded as the issue of two-class, and the feature subset was selected from the above feature set based on heuristic searching and filter pattern with an assessing function of class separability criterion. To select the feature subset based on optimal scalar feature, class separability measure of every scalar feature was calculated and sorted by descending on training set, and the frontal l scale features were assembled l feature subset to establish a Bayes-criterion based discrimination model (l=1,2,...,64or15) on training set. With increasing feature subset capacity of l, the error rate on training set kept decreasing while the error rate on validation set changed from decreasing to increasing, and the inflexion of which was used for the best subset capacity in practice. Experiment results showed that spectrum feature subset was selected at the best feature subset capacity of14, based on which the classification rate on prediction set was77.78%; geometric structure feature subset was selected at the best feature subset capacity of8, based on which the classification rate on prediction set was85.56%. To select the feature set based on floating searching,l feature subset (l=1,2,...,64or15) with maximal class separability measure was searched backward and forward to establish a Bayes-criterion based discrimination model on training set. With increasing feature subset capacity of l, the error rate on training set kept decreasing while the error rate on validation set changed from decreasing to increasing, and the inflexion of which was used for the best subset capacity in practice. Experiment results showed that the spectrum feature subset was selected at the best subset capacity of22, based on which the classification rate on prediction set was82.22%; the geometric structure feature subset was selected at the best subset capacity of8, based on which the classification rate on prediction set was85.56%.Ripeness and under-ripeness of raw cotton were considered as the issue of two-class, and the feature subset was selected from the above feature set based on exhaustive searching and wrapper pattern with an assessing function of error rate. All of l feature subset were used to establish Bayes-criterion based discrimination models on training set and l feature subset (l=1,2,...,64or15) with the minimum error rate on validation set was selected. With increasing feature subset capacity of l, the error rate on training set kept decreasing while the minimum error rate on validation set changed from decreasing to increasing, and the inflexion of which was used for the best subset capacity in practice. Experiment results showed that the spectrum feature subset was selected at the best subset capacity of15, based on which the classification rate on prediction set was83.33%; the geometric structure feature subset was selected at the best subset capacity of6, based on which the classification rate on prediction set was88.89%.3. The color space measuring cotton color was selected based on correlation analysis. Cotton color was measured in Hunter, Lab, HSI color space according to Chinese government grading standard of raw cotton; and correlation analysis showed that some correlation coefficients of cotton color are false in three color space. With adjusting the component of color space in gray level, all of correlation coefficient are true in HSI color space, which can be used to measure cotton color.The texture and shape feature set used to discriminate quality grade of raw cotton was extracted in spatial domain.12texture features, including mean, standard deviation, smoothness, the3rd-order moments, consistency, and entropy, were extracted based on the histogram of S and I in HSI color space to describe cotton color and impurity content, including rain rust, hoarfrost, stiff petal, stain, and rot boll.16shape features, including the number of cotton petals, global feature, three radial cut region features, five circumferential cut region features, two radial cut line features, three circumferential cut line features, and box-counting dimension, were extracted based on ratio method to describe cotton size and geometric structure distributing. A valid feature subset without more0.9correlation coefficient was selected based on the correlation of the average texture/shape feature and quality grade from1to7at the0.05level, including the3rd-order moment of I, mean of S, standard deviation of S, entropy of S, the number of cotton petals, global feature, the1st and4th circumferential cut region feature, the1st radial cut line feature, the3rd circumferential cut line feature, and box-counting dimension.Quality grade of raw cotton was regarded as the issue of seven-class, and the feature subset was selected from the above valid feature set based on10-fold cross-validation, heuristic searching and filter pattern with an assessing function of class separability criterion. To select the feature subset based on optimal scalar feature, class separability measure of every scalar feature was calculated and sorted by descending on every training set, and the frontal l scale features were assembled l feature subset to establish a Bayes-criterion based discrimination model (1=1,2,...,64or15) on every training set. With increasing feature subset capacity of l, the average error rate on10training sets kept decreasing while the average error rate on the10corresponding validation sets changed from decreasing to increasing, and the inflexion of which was used for the best subset capacity in practice. Experiment results showed that the10valid feature subsets were selected at the best subset capacity of7on10training sets respectively, and the fifth feature subset was selected with the minimum error rate on the corresponding validation set, based on which the classification rate on prediction set was74.71%. To select the feature set based on floating searching,l feature subset (l=1,2,...,11) with maximal class separability measure was searched backward and forward to establish a Bayes-criterion based discrimination model on every training set. With increasing feature subset capacity of l, the average error rate on10training sets kept decreasing while the average error rate on the10corresponding validation sets changed from decreasing to increasing, and the inflexion of which was used for the best subset capacity in practice. Experiment results showed that the10valid feature subsets were selected at the best subset capacity of6on10training sets respectively, and the ninth feature subset was selected with the minimum error rate on the corresponding validation set, based on which the classification rate on prediction set was70.11%.Quality grade of raw cotton was considered as the issue of seven-class, and the feature subset was selected from the above valid feature set based on10-fold cross-validation, exhaustive searching and wrapper pattern with an assessing function of error rate. All of l feature subset were used to establish Bayes-criterion based discrimination models on every training set and l feature subset (l=1,2,...,11) with the minimum error rate on the corresponding validation set was selected. With increasing feature subset capacity of l, the average error rate on10training sets kept decreasing while the average minimum error rate on the10corresponding validation sets changed from decreasing to increasing, and the inflexion of which was used for the best subset capacity in practice. Experiment results showed that the10valid feature subsets were selected at the best subset capacity of9on10training sets respectively, and the sixth feature subset was selected with the minimum error rate on the corresponding validation set, based on which the classification rate on prediction set was79.31%.4. The feature subset based on exhaustive searching and wrapper pattern is optimal, which can availably eliminate redundant features and identify key features with large computing cost and low speed. The feature subset based on heuristic searching and filter pattern is suboptimal, which may select redundant features with little computing cost and high speed. And an optimistic feature subset may be selected, the error rate of which on validation set was less than that on training set, possibly, prediction set less than validation set.10-fold cross-validation based discrimination model is more steady and reliable.In order to discriminate ripeness of raw cotton, the geometric structure feature subset was selected with a generalization precision of88.89%based on wrapper pattern and exhaustive searching finally, including the1st and2nd radial cut region feature, circumferential cut region feature (1st,3rd and4th), the2nd radial cut line feature, which revealed that the inner of ripe cotton is dissilient while outer is plump, and the inner of unripe cotton is close while outer is lank.In order to discriminate quality grade of raw cotton, the texture and shape feature subset was selected with a generalization precision of79.31%based on10-fold cross-validation, wrapper pattern and exhaustive searching finally, including the3rd-order moment of I, mean of S, standard deviation of S, and entropy of S, the number of cotton petals, global feature, the1st circumferential cut region feature, the1st radial cut line feature, and box-counting dimension. The feature subset revealed that high grade cotton petal is whiter, lower yellow, less impurity, more compact, thicker, and fluffier, and that the low grade cotton petal is more light-grey, higher yellow, more impurity, sparser, thinner, stiffer, much stain, and much rot boll. Raw cotton was usually misclassified to neighboring grade, and the higher and lower grade raw cotton were identified exactly. |
PDF Full Text Request