结合深度学习与图像处理的黄斑茧识别算法

Yellow Spotted Cocoon Recognition Algorithm CombinedIwith Deep Learning and Image Processing

剥茧缫丝是提取丝绸制品原料蚕丝的重要环节,在缫丝前需要对蚕茧进行筛选,剔除不合格的下茧。针对目前自动化识别上车茧和黄斑茧准确率低的问题,文中提出一种结合深度学习与图像处理技术的识别算法。引入空洞卷积改进YOLOv5s网络,利用改进后的网络对不同类别标签的蚕茧图片进行训练和预测;在此初识别基础上,对网络预测结果置信度小于70%的图片进行图像处理二次判别,在原始图片上根据网络预测的锚定框提取出蚕茧所在区域,经背景分割预处理后单独提取蚕茧HSV颜色模型中S通道图,在S通道图上分析蚕茧黄斑颜色特征,统计表面黄斑区域的面积占比和平均饱和度,并设定双阈值进行二次识别。经测试,该算法识别上车茧和黄斑茧的平均准确率达到94.94%,单张图片初识别加二次识别总时间为318.5 ms。

Cocoon stripping and reeling is an important step in extracting raw silk for silk products,and before silk reeling,unqualified cocoons must be removed.Aiming at improving the current low accuracy of automatic recognition between yellow spotted cocoon and reelable cocoon,in this paper,a recognition algorithm combining deep learning network model and image processing technology was proposed.The dilated convolution was introduced to improve the YOLOv5s network,so as to improve the network training of predicting cocoon pictures with different categories of labels.On the basis of the initial recognition results,images with a confidence coefficient of less than 70%were subjected to a secondary recognition.On the original image,the cocoon area was extracted according to the anchor box of the network prediction.After pretreatment of background segmentation,the S-channel image in the HsV color model of the cocoon was extracted separately.The color characteristics of the yellow spots were analyzed on the S-channel image,the area proportion and average saturation of yellow spots area were counted,and then,after the double thresholds were set,the secondary recognition was performed.The results of experiments showed that the average accuracy of the algorithm in differentiating the yellow spotted cocoon from the reelable cocoon was 94.94%,and the total time of the initial recognition and secondary recognition of a single picture was 318.5 ms.