| China is a big country of cotton production and consumption.Xinjiang is the major cotton producing province in China and its cotton production and quality play a key role in the healthy development of Chinese cotton textile processing industry.Drip irrigation under film is widely used in Xinjiang.This method can save irrigation water,improve fertilizer utilization rate and increase cotton output.However,during the process of mechanical harvesting,a large amount of residual mulch film is mixed into seed cotton.If the residual mulch film cannot be removed in time,it will seriously affect the subsequent cotton processing and reduce the quality of cotton textiles.It is difficult to identify the mulch film using traditional methods,because the film is colorless and transparent without fluorescent effect.This paper is a research on sorting system of film on seed cotton based on deep learning.novel algorithm based on variable-wise weighted stacked autoencoder and extreme learning machine was proposed.The main research contents and results are as follows:1.A set of film sorting system on seed cotton was designed,which included three parts: seed cotton opening device,film detection device and film removal device.The function of the seed cotton opening device is to break up the cotton lump for the convenience of subsequent detection and identification.The film detection device is composed of hyperspectral image acquisition module and seed cotton and film recognition algorithm module.Its function is to realize the recognition and classification of seed cotton and film.The function of the film removal device is to remove the identified film in time.2.This paper introduced the characteristics of hyperspectral image data and the common methods of dimensionality reduction and classification.The hyperspectral image(1000-2500nm)samples of seed cotton and film were collected and the spectral information of the samples was analyzed.The number of classification categories(4 categories: background,film on background,seed cotton,film on seed cotton)was preset.This paper described the basic principles of autoencoder and extreme learning machine,and proposed a variable-wise weighted autoencoder algorithm,which assigned different weights to each band,suppressed the impact of noise,maintained the multi-channel advantage of hyperspectral data,and extracted robust features.In order to enhance the performance of the network,the variable-wise weighted stacked autoencoder was constructed by cascading multiple variable-wise weighted autoencoder to extract data features.The last hidden layer neuron was used as the input of the classifier.The classifier used extreme learning machine.At the same time,the gray wolf optimization algorithm was used to select the number of hidden layer nodes and parameters of extreme learning machine to improve the classification accuracy.All pixels in the hyperspectral image were classified,and the corresponding probability matrix can be obtained for each category.The probability matrix of each category was processed by morphological method to eliminate some false recognition areas and noise points,and then the pixel label could be determined by the maximum probability of the category.Finally,the results were merged into film and non-film two classes.,which was the final result.The experimental results show that the averaged recognition rate of the proposed algorithm is 95.58%.3.The algorithm was integrated into the sorting system,which was configured with NVIDIA GTX1060 GPU.The machine was tested in two companies in Shandong and put into production in Xinjiang.The algorithm run in Keras framework and was implemented by CUDA and CUDNN.After repeated tests,when the output is 3t/h,the film selection rate of the sorting system reaches 95%,which meets the design requirements.Therefore,the algorithm proposed in this paper can be used in actual production. |
