摘要
针对小麦不完善粒识别中传统图像处理方法需要复杂的特征提取且识别效果不佳的问题,设计并实现基于CNN深度模型的小麦不完善粒识别方法。建立图像数据库Wheat Image,并结合空间金字塔池化理论构建CNN网络模型,接着对样本集进行扩展,以提高模型泛化能力,设计双面识别方案并完成对小麦完善粒、破碎粒和病斑粒的识别。所提出的方法相对于传统的图像处理识别方法,识别率提高15个百分点;相对于常规CNN模型,识别率提高5%;对于引入噪声以及亮度改变的图像,识别率也达到90%以上;设计的双面识别方案有效地降低了识别的错误率。提出的方法不仅避免复杂的特征提取步骤,而且有效地提升麦粒识别率,对小麦的智能检测识别具有重要意义。
In identification of the unsound kernels of wheat, traditional image processing method needs complex feature extraction and can lead to poor recognition. Therefore, proposes a novel identification method based on CNN deep model. Establishes an image database named Whea- thnage, and builds CNN network model combined with the theory of space pyramid pool. Then, the dataset is extended to improve the mod- el generalization ability. Finally, designs the double-sided recognition scheme to complete the recognition of sound kernels, broken kernels and spotted kernels. This method improves recognition rate by about 15% compared to traditional method of image processing. And it im- proves recognition rate by about 5% compared to the conventional model of CNN. For noisy images and images that the brightness is changed, the recognition rate of the proposed method reaches more than 90%. The scheme of double identification effectively reduces recog- nition error rate. The proposed method not only avoids complex feature extraction, but also effectively improves recognition rate, which is of great significance to the intelligent detection and identification of wheat.
出处
《现代计算机》
2017年第24期9-14,共6页
Modern Computer
基金
国家自然科学基金委员会和中国工程物理研究院联合基金(No.11176018)
成都市科技惠民项目(No.2015-HM01-00293-SF)
特殊环境机器人技术四川省重点实验室开放基金(No.14zxtk03)
