Abstract:

Objective It has been shown that image processing methods can not clearly acquire image characteristics of foreign fibers in cotton layers. In order to solve the problem associated to conventional image processing methods, this paper proposed a classification and identification method for foreign fibers in cotton layers based on near-infrared (NIR) spectroscopy and residual neural networks (ResNet).

Method In this study, 500 groups of foreign fibers spectral data were collected by experiments, including five types of foreign fibers. The spectral collection instrument was a UH4150 spectrophotometer. Savitzky-Golay method was adopted to smooth the spectral data, and F-test and LightGBM classification algorithm was adopted to determine the optimal feature wavelength. The optimal spectral data were converted into Garmian angular summation fields (GASF) images by the Garmian angular field (GAF) method, which preserved the temporal sequences between wavelength sequences. Eventually, the ResNet model was constructed. The GASF images were used as training samples to train the ResNet model.

Results The foreign fibers' spectral data was smoother than the original spectrum by the Savitzky-Golay method. Noisy data at both ends of the spectrum and near the peaks of functional groups were eliminated (Fig.2). After F-test and LightGBM classification algorithm wavelength optimization, 75 optimal wavelengths were selected. When the number of wavelengths was greater than 200, important information was deleted from the foreign fibers' spectral data (Fig.3(a)). When the number of wavelengths was 75, the optimal performance of the optimized model was the best, and the accuracy reached 98.99% (Fig.3(b)). The accuracy of applying the GASF image to the ResNet model is 99.69%(Fig.7(a)). The loss of the training set showed a sharp drop for the first 50 iterations (Figs.7 (b) and (c)). When the number of iterations reached 70, the training set started to converge. When the number of epochs reached 200, the training set tended to be stable. The classification accuracy of gray-scale and time-frequency images was lower than 99.00%, lower than the recognition accuracy of GASF images(Fig.8). The ResNet model improved classification performance compared with machine learning classification models. Compared with the K-nearest-neighbor (KNN) and decision tree (DT), accuracy increased by 6.67% and 7.60%, and recall increased by 6.91% and 7.09%. Compared with the artificial neural network's multi-layer perceptron (MLP), the accuracy and recall increased by 1.22% and 1.36%, respectively. All the classification performances of rope and feather on the ResNet model reached 100%, indicating excellent classification of these two types of foreign fibers. There were no false positives or missed inspections. Only two misjudged cases were found in 640 foreign fibers samples in the test set, the first being the chemical fibers were wrongly judged as PP yarn. Second, the chemical fiber was wrongly identified as plastic bag.

Conclusion The classification and identification model based on GASF and ResNet improved the feature extraction performance of foreign fibers' near infra-red spectra data. This method can effectively identify the foreign fibers in the cotton layer under a complex environment, with a identification accuracy of 99.69%. The classification and identification method of foreign fibers based on spectra combined with the convolution neural network (CNN) provides new research ideas for classifying foreign fibers in a complex environment. In addition, the method provides technical support for developing the sorting device of foreign fibers. Future research will be further extended to cotton quality evaluation fields, such as content detection of foreign fibers.

Key words: cotton, foreign fiber detection, near-infrared spectrum, residual neural network, spectral data, image detection