Identification Of Plastic Films In Ginned Cotton Based On Hyperspectral Imaging Technique

Plastic films are one of the most common foreign matter in cotton whose texture is light, color is transparent or similar to cotton and are easily attached to the cotton fiber which make the identification rate of traditional detection methods low and lead to the fact that identification of plastic films has been a tough problem for cotton industry that needs to be solved. Therefore, the plastic films in cotton were chosen as the research object and the visible-near infrared hyperspectral reflectance images of them were acquired. Combining with spectral analysis, image processing and machine learning techniques, the identification methods for plastic films in cotton were studied. The main research work were as follows:(1) Optimal bands selection methods of hyperspectral images of plastic films in cotton. Through the analysis of the spectral data and image data of plastic films and cotton, a characteristic wavelength selection method based on the difference spectra analysis and principal component analysis was proposed. Firstly, the difference spectra between the plastic films and cotton were analyzed and two wavelengths with the maximum spectral difference were selected as characteristic wavelengths; then, principal component analysis was conducted on the hyperspectral images of plastic films and the first three principal components images were chosen to analyze their weighing coefficients. A total of 6 optimal characteristic wavelengths were selected. In addition, the spectral data of plastic films were analyzed by using partial least squares regression and 4 characteristic wavelengths were chosen. The results showed that the wavelengths selected by the two methods could be effectively used for the detection of plastic films in cotton.(2) Multi-band image fusion and segmentation methods of plastic films in cotton. The study on multi-band image fusion and segmentation methods of hyperspectral images were carried out and two image segmentation methods of hyperspectral images of plastic films based on multi-band image fusion methods were proposed. â‘  The images corresponding to the 6 wavelengths selected by difference spectra analysis and principal component analysis were extracted and calculated to get the mean images. Gaussian low-pass filtering was then used for pre-processing, a fixed threshold method and an improved iteration method were used for image segmentation, respectively. The results showed that the two methods got similar segmentation results and the Area Overlap Measure (AOM) value of the improved iteration method was 0.6439 which was higher than that of the fixed threshold method although it required more time. â‘¡ The images corresponding to the 4 wavelengths selected by partial least squares regression were extracted and fused using arithmetic operation methods, and then image segmentation, median filtering, small targets removals and other operations were performed. Furthermore, Gaussian low-pass filtering, the improved iteration method, median filtering, small targets removals were performed on the first principal component image, the first minimum noise fraction image of the 4 selected band images and the 4 single-band images to get them enhanced and segmented to compare with the results of the proposed methods. The results showed that the AOM value of the proposed method was 0.6636 and was higher than those of the other methods, which indicated that the proposed method achieved more accurate segmentation of plastic films.(3) Feature extraction and feature selection methods of plastic films in cotton. The shape features and texture features of plastic films were extracted and analyzed, and an improved minimal redundancy maximal relevance method for feature selection was proposed. Firstly,7 Hu moment invariants (shape features),6 GLCM features (texture features),40 mean values of energy images after Gabor transformation (texture features) of the plastic films were extracted to form a 53-dimensional feature set and the data were normalized. Then, the minimal redundancy maximal relevance method and the improved minimal redundancy maximal relevance method were used to analyze the feature set, respectively. The results showed that the optimal features subset of the first 5 features selected by the improved minimal redundancy maximal relevance method got the classification rate of 97.71%, achieved the dimension reduction of high-dimensional feature set and improved the classification accuracy and speed of the plastic films.(4) Intelligent identification methods of plastic films in cotton. By studying the parameters optimization methods of different particle swarm algorithms, a new identification method based on improved particle swarm optimization (PSO) and support vector machine was proposed. Firstly, penalty parameter and kernel function parameter were calculated using particle swarm optimization algorithm and 3 improved particle swarm optimization algorithms and were then used for identification with support vector machine whose kernel function was radial basis function. Then the minimal penalty parameter obtained by the improved PSO was used for further optimization while the kernel function parameter was kept unchanged. The results showed that an optimal set of penalty parameter and kernel function parameter was obtained with the identification accuracy of train sets and test sets were both 100% which indicated that the classification performance was improved.