基于深度学习的籽棉中异性纤维检测方法

针对籽棉图像阴影多、常规图像处理方法难于识别的问题，以去除棉叶、棉壳等有机杂物的籽棉为样本，将不同颜色、形状、尺寸的12种常见异性纤维和籽棉样本随机地分布在运转中的传送带上，采用线扫描相机获得发光二极管（LED）照明的籽棉图像520张，“LED+线激光”双光源照明的籽棉图像1 148张。然后采用一种由13个卷积层、13个采样层和4个池化层构成的Faster RCNN深度学习人工神经网络，对 2 种成像方法获得的籽棉图像进行基于人工智能的网络训练，再进行异性纤维检测验证。实验数据表明，LED照明和“LED+线激光”双光源照明条件下，籽棉图像中的异性纤维的检出率分别达到了90.3%和86.7%，特别是LED照明条件下对白色异性纤维进行识别，其识别率由5.9%提升到了90.3%。

Detecting method of foreign fibers in seed cotton based on deep-learning

In order todetect foreign fibers in seed cotton with heavy shadows, seed cotton samples without cotton shell and leafs, and 12 types of foreign fibers with different colors, shapes and sizes were randomly distributed on amoving convey surface. And then, 520 seed cotton images were obtained under the illuminations of light emitting diode (LED) and 1 148 images were obtained under the illuminations of double light source of LED+ linear laser by a color line-scan camera. Then Faster RCNN deep-learning neural networkscomposedof 13 convolutional layers, 13 sampling layers and 4 pooling layers were constructed. After training, the neural networks were used for detecting foreign fibers in the two types of seed cotton images respectively. The experimental results indicated that the detecting rates of the targets in the images under the illumination of LED and LED + linear laser are 90.3% and 86.7%, respectively, by the Faster RCNN. Especially,the detecting rate of white color foreign fibersincrease from 5.9% to 90.3%.