Research On Seed Cotton Foreign Fiber Sorting Recognition Algorithm Based On Deep Learning

China is a big consumer of cotton,and Xinjiang cotton area is the largest cotton producing area in China.Its cotton production and quality play a vital role in the processing,production and development of China's cotton textile industry.In recent years,with the continuous increase of labor costs,the mechanical harvesting of cotton has been promoted on a large scale,so that a large amount of impurities including the mulch film are mixed into the seed cotton.Especially the mulch film,which accounts for up to 90% of the impurities,the traditional conventional method cannot effectively detect such mulch film,which greatly affects the quality of cotton and seriously affects the quality of textiles,and brings huge losses to the domestic cotton planting industry.This thesis is a research on the research of seed cotton foreign fiber sorting recognition algorithm based on deep learning,a seed cotton film recognition algorithm based on variable weighted convolutional neural network was proposed,and the seed cotton foreign fiber recognition algorithm based on target tracking was studied.The main research contents and results are as follows:(1)A set of seed cotton film processing system was designed,including seed cotton opening device,seed cotton film recognition and rejection system.The identification and rejection system mainly includes hyperspectral image acquisition module and deep learning algorithm identification module.(2)A seed cotton mulching recognition algorithm based on variable weighted convolutional neural network is proposed for the mulch film which is difficult to sort in seed cotton.For the hyperspectral image of the collected seed cotton mulch,the characteristic band is extracted first by PCA algorithm.After preprocessing,200 bands are reserved,and the one-dimensional spectral vector of each pixel is input into the variable weighted convolutional neural network for identification.The variable weighted convolutional neural network proposed in this paper includes a module called an important factor block,which is derived from the auto-encoder,and it adds a weight to each element of the cell vector,so that the wavelength of the larger weight is more affected in the subsequent network,and the influence of the smaller weight wavelength in the subsequent network is weakened.The experimental results show that the classification accuracy of the seed cotton mulching algorithm based on variable weighted convolution neural network is higher than that of the traditional algorithm.The overall accuracy and average accuracy of two test images are 98.41% and 98.27%,98.93% and 98.81%.(3)For general seed cotton fiber such as color line,hair,cloth strip,etc.,this paperattempts to identify the seed cotton foreign fiber based on target tracking,including Faster R-CNN and SSD.The Faster R-CNN processing of the input image consists of two parts.One is to use the regional suggestion network RPN to predict the foreign fiber position and give a suggestion box,where RPN is a full convolutional neural network;the other one is to use ZFNet to extract the image features of the suggestion box and classified.RPN shares the convolutional layer with ZFNet.Based on the previously fed convolutional neural network,the SSD network can specify the bounding box size and remove the extra borders by non-maximum value suppression to obtain the classification result.Test images were tested using Faster R-CNN,and the detection rate of all foreign fibers was 83.93%.The test images were tested using SSD network,and the detection accuracy of all foreign fiber types was73.21%.