Research On Profile Fiber Recognition Based On Genetic Algorithm And BP Neural Network.

The profile fiber existing in cotton affects the quality of textiles, which will reduce the economic benefits of cotton textile enterprises. For conventional manual picking method, the efficiency is very low, and has the high probability of missed and wrong detection. With the development of computer vision, the technology of automatic detection has drawn widely attention. In this thesis, we research on the detection method of profile fiber based on image processing technology, and propose a detection method based on genetic algorithm and BP neural network.Firstly, we apply histogram equalization and media filtering method into the fiber image to reduce the effect of noise. The effects of several segmentation methods are analyzed, and we select the morphological method as the segmentation algorithm in this thesis.Secondly, we choose the feature with the small distance within the class and the larger distance between the class based on analyzing the character of fiber images. In this thesis, we adopt the gray-level co-occurrence matrix to depict the texture of fiber image because of this capability of blending spatial interaction with gray-level distribution. For the co-occurrence matrix, we calculate its entropy, energy, moment of inertia and relevant, and get a feature vector with 8-dimensions by calculating mean and variance of the four parameters. For the second feature, we select Gabor wavelet filter sets with six scales and four directions to filter the image to get 24 images, and get the feature with 48-dimensions by calculating the mean and variance of the 24 Gabor images. According to the calculating the variance of the same feature between different images, we reduce the 48-dimensions features into 36-dimensions.In the end, we propose an improved BP neural network based on genetic algorithm, and adopt the method to classify the different kinds of the profile fiber. Firstly, we determine the number of neurons of BP neural network at input layer based on the dimension numbers of the feature. Then, based on the number of the kinds of profile fibers, we get the number of neurons at output layer. After selecting the neural network structure, we code the random weights and thresholds. And the improved BP neural network is achieved by optimizing the initial weights and thresholds using generetic algorithm. We train the improved neural network using training sets, and classify the test sets based on the trained neural network.Experimental results show that the proposed method can efficiently detect and classify the profile fiber. And the proposed method has a good prospect.