Hyperspectral image segmentation method for plastic films in ginned cotton

Traditional vision detection methods are not suitable for plastic films mingled in ginned cotton, for their color is transparent or very similar to cotton. As an emerging technique, hyperspectral imaging integrates spectroscopy and traditional imaging, can acquire spectral and spatial information at the same time. Therefore, hyperspectral reflectance images of plastic films were captured using hyperspectral imaging technique and a hyperspectral image segmentation method for plastic films by the combination of spectrum analysis and image processing technology was proposed. Eight regions of interest(ROIs) with different sizes in different parts of each sample were randomly selected to extract spectral data. The mean spectra of each sample were obtained by calculating the average spectra of all the 8 ROIs. The noisy spectral data at the two ends of the spectral curves were removed and 750 bands of spectral data in 450.8~1034.9 nm range were selected. In order to decrease noise disturbance, Savitzky-Golay(SG) polynomial smoothing was used to eliminate random noise, and standard normal variate(SNV) transformation was used to correct the errors caused by spectrum scattering. The preprocessed spectral data matrix was selected as independent variables to establish the partial least squares regression model. Then, the regression model was used to conduct partial least squares regression analysis, and the first three loading vectors of the loading matrix were chosen to plot the curves. Significant difference in the peaks and valleys of the three loading vectors from zero was observed, therefore, the wavelengths corresponding to the peaks and valleys could be considered as effective wavelengths. Then 4 wavelengths at 560.3, 673.9, 716.9 and 798.8 nm were selected as the effective wavelengths. Afterwards, arithmetic operations were used for image fusion. Images at the 4 wavelengths were extracted and the addition of three images at 673.9, 716.9 and 798.8 nm was performed. The result showed that the main part of the plastic films was retained through a variety of image fusion methods. Hence, the fused image was used for segmentation by Otsu’s method and then median filtering was performed on the segmentation results to reduce noise. In order to obtain the edge information of plastic films, a subtraction of the doubled images at 560.3 nm from the aforementioned fused image was adopted to obtain the fused images. Otsu’s method and median filtering were performed again to get the segmentation results and remove noise. Finally, the results obtained with the abovementioned two fusion methods were integrated again and small targets were removed to get the final results. Segmentation results showed that the plastic films were well segmented and the edge details were complete. To better reflect the effectiveness of the used image fusion method, principal components analysis(PCA) and minimum noise fraction(MNF) were conducted on the 4 images at the effective wavelengths and Otsu’s method was used to segment the first principal component(PC1) image and first minimum noise fraction(MNF1) image. Then the median filtering and small targets removal were performed. Observation revealed that the over-segmentation of PC1 and MNF1 images was serious and many background areas were divided into target areas, which indicated that the proposed image fusion method was better than both PCA and MNF. To validate the segmentation performance of the proposed method, area overlap measure(AOM), misclassified error(ME) and recognition accuracy were used to evaluate the segmentation effects objectively. The results showed that the proposed method got the best AOM value of 0.6636 and ME value of 0.4226, with the accuracy reached 91.07%. This study reveals that the proposed method is able to well segment the images of plastic films and can lay a good foundation for the subsequent feature extraction and automatic recognition.