Research On Methods Of Tobacco Foreign Material Detection And Tobacco Leaves Grading Based On Machine Vision

Computer vision technology provides a feasible way for the automation of tobacco industry processing. The technologies of tobacco foreign material detection and tobacco leaves automatic grading are paid attention by numerous researchers, and the researchers already have done some study. To solve the problem of lower detection accuracy and lower grading accuracy in current methods, it is important in theory and practice to study more effective methods of tobacco foreign material detection and tobacco leaves grading. The methods of tobacco images processing are mainly studied in this paper, such as foreign material detection for tobacco, denoising for tobacco leaf images, classification for tobacco leaves of normal group and abnormal group, grading for tobacco leaves of normal group. The mainly work of this paper is as follows:Firstly, a threshold segmentation algorithm for noisy images is proposed based on non-local similarity cross entropy. And this algorithm is directly applied to noisy tobacco images for foreign material detection. Threshold vector group is figured out adaptively according to gray levels of the noisy image in this algorithm. The threshold vector in threshold vector group is used in turn to segment noisy images. Non-local similarity between threshold vector and eight neighborhood vector of each pixel is used to express the probability distribution. According to this method, the probability distributions of original image and segmented image are figured out respectively. Noisy images are segmented by the optimal threshold vector which is selected to detect tobacco foreign material according to the rule of cross entropy. After a lot of experiments, compared with many threshold segmentation algorithms based on entropy, the algorithm proposed in this paper has obviously advantage of noise immunity and strong robustness. And it is suitable for detecting the foreign material which is great difference in gray with tobacco.Then, a threshold segmentation algorithm of bidirectional recursion is studied based on improved gray entropy. And this algorithm is applied to the result of tobacco images preprocessing to detect foreign material. Corresponding the one-dimensional and two-dimensional histograms, the representation method of image probability distribution is presented respectively in this improved gray entropy. Tsallis entropies are calculated with different thresholds in this probability distribution. Bidirectional recursion algorithm is used to speed up the computational efficiency. The threshold corresponding the maximum Tsallis entropy is chosen to segment the image for foreign material detection. Experimental results show that, compared with the threshold segmentation algorithms of two dimensions gray entropy, such as decomposition algorithms based, recursive algorithms based, optimization algorithms based. The algorithm presented can better represent the gray consistency in the same class, the result of foreign material detection is more accurately.Next, a non-local means image denoising algorithm is proposed based on clustering by steering kernel and adaptive search windows in this paper. This algorithm is applied to tobacco leaves for denoising. Fuzzy c-means clustering algorithm based on steering kernel is used to prescreen and classify similar windows. The size of search windows corresponding to each pixel is calculated according to categories of similar windows. The number of similar windows with higher similarity is guaranteed. Image denoising of non-local means based on adaptive search windows is carried out for each category. After a lot of experiments of denoising for tobacco leaves images, compared with the three improved non-local means algorithms such as the algorithm based on Zernike moment, the algorithm based on principal neighborhood dictionaries, and the algorithm based on prescreening of mean-variance, the proposed improved non-local means algorithm has better denoising effect for the tobacco leaves images with strong noise or irregular texture. The textures and edges of tobacco leaves are better preserved.Subsequently, a classification algorithm for tobacco leaves of normal group and abnormal group is achieved based on multi-features and wavelet support vector machine. Green rate and red roasted rate as two color feature are defined in the HSV spaces of tobacco leaves. Texture features are extracted by Gabor wavelet. After normalized, these features are inputted into wavelet support vector machine for training. Each tobacco leaf image is classified by trained wavelet support vector machine. Experimental results show that, compared with the classification algorithms for tobacco leaves of normal group and abnormal group, such as the algorithm based on fuzzy inference, the algorithm based on back propagation neural network, the algorithm based on support vector machine, this algorithm achieves better classification result for tobacco leaves. Error between normal group and abnormal group is lower. And correct rate of classification is high.Finally, a grading algorithm for tobacco leaves of normal group is explored based on convolutional neural network and color recognition. Color tobacco leaf images are transformed to gray images. And the gray images are directly inputted into convolutional neural network for parts recognition of normal tobacco leaves. At the same time, RGB spaces of tobacco leaf images is transformed to HSV spaces. Threshold intervals of red, orange and yellow are defined in channel H. And the proportional distribution of these three colors is counted to achieve color recognition of tobacco leaves at the same part. Experimental results show that, compared with the grading algorithms for tobacco leaves of normal group, such as the algorithm based on fuzzy inference, the algorithm based on generalized regression neural network, the algorithm based on support vector machine, this algorithm describes normal tobacco leaf features of different parts better. The algorithm of color recognition is stabilized. Tobacco leaves of normal group are graded better.