小麦籽粒多角度图片库构建

Construction of multi-angle picture library of wheat grain

小麦是我国的重要粮食作物,小麦籽粒识别和分类是谷物质量评估的重要任务.小麦在视觉上的表现因为角度不同而差别较大,为精确提取小麦特征,在已有数据集构建方法基础上采用同一小麦籽粒多幅图片,增加了分角度处理,基于小麦籽粒腹沟向上、腹沟向下和腹沟朝前3个角度构建了分角度小麦籽粒品种图片库.首先采集黄淮麦区种植面积较大的6个小麦品种各1000粒,分别拍摄每粒小麦,对所有图片进行自动更名和预处理得到包含18000张图片的分角度小麦籽粒分类数据库.为验证分角度数据采集的合理性,采用VGG16网络进行小麦籽粒识别,实验结果中角度统一后识别效果较好,特别腹沟朝前角度性能优于其他两个角度,数据增强后其准确率可达98.7%,优于角度混合数据集的识别准确率.实验结果表明,构建的多幅同一小麦籽粒图片分角度小麦籽粒分类数据库有助于分类模型更准确地提取小麦籽粒特征,避免了已有数据集中采集角度不统一造成的特征干扰问题,以较少数据量获得较高识别率,提高了品种识别的兼容性和准确性.

Wheat is an important grain crop in China.Wheat grain recognition and classification is an important task of grain quality evaluation.The visual performance of wheat varies greatly due to different angles.In order to accurately extract the characteristics of wheat,based on the existing data set construction methods,this paper uses multiple pictures of the same wheat grain,adds sub angle processing,and constructs a sub angle wheat grain variety picture library based on the three angles of wheat grain belly ditch up,belly ditch down and belly ditch forward.Firstly,1000 grains of 6 wheat varieties with large planting area in Huang Huai wheat area were collected,and each grain of wheat was taken separately.All pictures were automatically renamed and preprocessed to obtain a sub angle wheat grain classification database containing 18000 pictures.In order to verify the rationality of sub angle data collection,vgg16 network is used for wheat grain recognition.In the experimental results,the recognition effect is better after the angle is unified.In particular,the performance of ventral groove forward angle is better than the other two angles.After data enhancement,its accuracy can reach 98.7%,which is better than that of angle mixed data set.The experimental results show that the multi-angle wheat grain classification database constructed in this paper helps the classification model to extract wheat grain features more accurately,avoids the feature interference caused by the inconsistent collection angle of the existing data sets,obtains a higher recognition rate with less data,and improves the compatibility and accuracy of variety recognition.