Detection Of Pseudo-Foreign Fibers Based On Computer Vision

During cotton planting, harvesting, storing, transferring, and producing, there are quite a few trashes that mixed into the fibers of lint cotton, which include botanical trashes such as cotton leaves, stems, grasses, and seed crumbs, named pseudo-foreign fibers or fake-foreign fibers, and harmful trashes such as Polypropylene wires, plastic films, mulch films and so on, named foreign fibers. Although the quantity of foreign fibers in cotton are not very much, they seriously affect the grade of cotton and the quality of afterwards textile product, and must be strictly culled in the textile factory. In real production environment, after several opening procedures, foreign fibers are always shattered into small pieces, similar in color, shape and texture to the large quantity of pseudo-foreign fibers in cotton. Thus in order to solve the hard problem of online detection of foreign fibers, and to improve the detection accuracy, a method of pseudo-foreign fibers detection based on Computer Vision was proposed in this dissertation, and the research works are described as follows:[1] The presegment enhance algorithms of lint images acquired from both the first and second experimental platform were studied, a method based on Variational Retinex for the second platform was proposed and compared with other algorithms such as Histogram Equalization, Wavelet based Normalization, Homomorphic Filtering, Single Scale Retinex and Multi-Scale Retinex. The experiment result showed that the method proposed are more suitable for the enhancement tasks of the second platform images of pseudo-foreign fibers, which can make the target more clearer and can facilitate the afterward works of image segmentation. A method combined the Laplace edge shapen algorithm and the improved Median Filter was proposed to enhance images from the first experimental platform, and the results showed that this method can strengthen the pseudo-foreign fiber targets in the first platform images effectively and efficiently which can meet the requirement of online image processes.[2] The characteristic of lint cotton images came from the first platform were studied, and a blocking and segmentation algorithm suitable for extraction of pseudo-foreign fiber targets was proposed. The cotton image was segmented averagely into thirty six small blocks at first, then the histogram of each block was examed for whether exist foreign fibers and pseudo-foreign fibers or not, if the answer is "yes", Otsu’s threatholding algorithm will be applied to segment the block and the Morphological Closing operator will be used for the post-segment process, and if the answer is "no", the block is labeled as background directly, at last all the processed thirty six blocks were merged into the segmented image, and the pseudo-foreign fiber targets were extracted. Compared with the method of using Otsu’s threathold directly to the whole image, Sobel edge detection and Robert algorithm, our method surpass them in both effectiveness and efficiency.[3] The feature extraction methods of pseudo-foreign fibers are studied. As foreign fibers, the five color features:Mean of R channel, Mean of G channel, Mean of B channel, Mean of average RGB, Standard deviation of average RGB are calculated, the seven shape features:Shape factor, Aspect ratio of external rectangle, Extent, Solidity, Eccentricity, Sphericity, and Euler number are also acquired, the nine texture features:Histogram based average brightness, Average contrast, Smoothness, Third moment, Consistency, Entropy, and Gray Level Co-occurrence Matrix (GLCM) based Angular second moment, Contrast and Entropy are computed too. In addition, ten color features based on HSV and Lab spaces are calculated, they are Mean of H channel, Mean of S channel, Mean of V channel, Mean of average HSV, Standard deviation of average HSV, Mean of L channel, Mean of a channel, Mean of b channel, Mean of average Lab and Standard deviation of average Lab, four rotation invariant texture features based on statistics of Local Binary Pattern Histogram Fourier transform (LBP-HF):Mean of fvl, Standard deviation of fvl, Mean of fv2and Standard deviation of fv2are computed and added into the feature set of pseudo-foreign fibers too.[4] Classification of foreign fibers and pseudo-foreign fibers was researched. The five color, seven shape and nine texture features of pseudo-foreign fibers were append to those of foreign fibers, Back Propagation (BP) Neural Network, Linear Kernel Support Vector Machine (SVM) and Radial Basis Function (RBF) Kernel Support Vector Machine (SVM) classifiers were applied to the whole lint trashes to classify their categories.[5]Clustering of pseudo-foreign fibers are researched, Simple K Means (SKM) method based on distance between instances and the Probability based on Expectation Maximize (EM) clustering algorithm are studied and applied to the fifteen color, seven shape and thirteen texture features of pseudo-foreign fibers, then Support Vector Machine classifier is used to test the two clustering method.[6] In order to testify the method mentioned above, two platforms are designed and implemented, one is based on Automatic Visual Inspection (AVI) and the other is based on Pneumatic Delivery and Computer Vision (CV). The hardware system of the first platform includes:lint input conveyer, lint layer generator, output roller, image acquisition system, lint layer collector and industry personal computer. The hardware system of the second platform includes:the blower, open machine, lint delivery passage, lint detection passage, image acquisition system, lint collector and industry personal computer. Both the first and the second platform’s software system includes:parameter setting, image acquisition, image processing, image analysis and report generation. Experiments on the platform indicate that the method for detection of pseudo-foreign fibers is right and effective.