Authors

Abstract

Daylily is widely used in medicine and diet therapy. In order to prolong the preservation period of daylily and make better use of its edible value, most of the daylily on the market are dried vegetables. Aiming at the problems of small size of dried daylily, similar color and texture between dried daylily, and difficulty in grading, this study proposes a method for grading dried daylily based on SSD. In the backbone feature extraction stage, the original backbone network VGG16 is replaced with the residual network model ResNet50 to realize the feature extraction of dried daylily. ResNet50 can deepen the network better and is more suitable for dried daylily feature extraction. Secondly, a feature fusion layer is added to improve the problem of insufficient utilization of shallow features in SSD network, which is more suitable for detail detection and improves the accuracy of dried daylily grading. Finally, the input image size is selected [512,512] to increase the image pixels, so that the network can capture more details of the dried daylily to improve the detection accuracy. The results show that the grading precision of the improved SSD algorithm is significantly improved compared with the traditional SSD, and the mean average precision is increased by 4.17%. At the same time, the same data set was used to test on the YOLOv5 model. Compared with YOLOv5s, YOLOv5s-CA and YOLOv5s-CBAM, the mean average precision was increased by 18.32%, 21.82% and 22.02% respectively, which further verified the precision and feasibility of the method and provided effective technical support for the grading of dried daylily.

Abstract in Chinese

黄花菜营养成分丰富，具有很高的食用和药用价值。鲜黄花菜因含有多种生物碱不宜多食，为了更好地利用其食用价值以及延长黄花菜的保存期，市面上的黄花菜大多是干菜。针对干制黄花菜体积小，黄花菜之间颜色和纹理相似，分级困难等问题，该研究提出了一种基于SSD的干制黄花菜等级分级的方法。该方法以SSD算法为基础，在主干特征提取阶段，将原主干网络VGG16替换为残差网络模型ResNet50，实现对干制黄花菜的特征提取。ResNet50可以更好地深化网络，更适合于干制黄花菜细节特征提取。其次，添加了特征融合层，改善了SSD网络中浅层特征利用不足的问题，更适合细节检测，提高了干制黄花菜分级的精度。最后，输入图像尺寸选取[512,512]，提高图像像素，使网络可以更好地捕捉干制黄花菜的细节信息，以提升检测精度。结果表明，改进后的SSD算法与传统SSD对比，分类精确度有明显的提升，平均精确度达到97.52%，对比原SSD算法提高了4.17%。同时，利用相同数据集在YOLOv5模型上进行试验，对比YOLOv5s、YOLOv5s-CA、YOLOv5s-CBAM，平均精确度分别提升18.32%、21.82%、22.02%，进一步验证了该方法的准确性和可行性，可以为干制黄花菜分级提供有效的技术支持。