基于多特征融合和深度置信网络的稻田苗期杂草识别

杂草的准确识别是田间杂草精准防控管理的前提,机器视觉技术是实现杂草准确识别的有效手段。该文以水稻苗期杂草为研究对象,采集稻田自然背景下和不同光照条件下的6种杂草图像共928幅,包括空心莲子草、丁香蓼、鳢肠、野慈姑、稗草和千金子。采用1.1G-R颜色因子将杂草RGB图像进行灰度化,选择自动阈值自动分割得到杂草前景二值图像,通过腐蚀膨胀形态学操作进行叶片内部孔洞填充,应用面积滤波去除其他干扰目标,最后将杂草二值图像与RGB图像进行掩膜运算得到去除背景的杂草图像;提取杂草图像的颜色特征、形状特征和纹理特征共101维特征,并对其进行归一化处理。在双隐含层和单隐含层的深度置信网络(deep belief networks,DBN)结构基础上,对DBN隐含层节点数选择方法进行研究。针对双隐含层DBN节点数,选择恒值型、升值型和降值型3种节点组合方式进行优化研究,当网络结构为101-210-55-6时杂草识别率为83.55%;通过对单隐含层节点参数优化得到网络结构为101-200-6时杂草识别率达到91.13%。以同一测试样本的运行时间值作为模型的测试时间对3种不同模型进行耗时测试,SVM模型、BP模型和DBN模型测试结果分别为0.029 7、0.030 6和0.034 1 s,试验结果表明基于多特征融合的DBN模型的识别精度最高,且耗时较其他2种模型相差不大,可满足实时检测的速度要求,所以在实际应用中应优先选择基于多特征融合的DBN模型。该研究可为稻田杂草识别与药剂选择性喷施提供参考。

Recognition of weeds at seedling stage in paddy fields using multi-feature fusion and deep belief networks

Abstract: Weed identification was the key to the site-specific weed management in the field. The machine vision method was adopted to realize automatic and rapid detection of weeds. This paper selected 6 weed species in paddy fields, including Alternanthera philoxeroides, Eclipta prostrata, Ludwigia adscendens, Sagittaria trifolia, Echinochloa crus-galli, and Leptochloa chinensis, which were captured in early growth stages with natural background and variable illumination. A total of 928 images were taken. The Alternanthera philoxeroides, Eclipta prostrata, and Ludwigia adscendens were dicotyledonous weeds which had large heart-shaped opposite leaves, and the other 3 weed species were monocotyledonous weeds which had narrow leaves. The image was 640×480 pixels and only a single seedling of weed was in the scene, and the acquisition format was color images of RGB (red, green, blue). The component with 1.1G-R was applied to gray level transformation of original RGB images. The OTSU adaptive segmentation method was adopted to realize the image segmentation of grayscale image. The morphological operation was used to fill vacancies in weed images. The noises and small target were eliminated based on area-reconstruction operator. The background was removed by masking algorithm between binary image and original RGB images. The 101-dimensional features were extracted from the foreground image of weed, including color, shape and texture feature. The color feature was composed of the first, second and third moments, the shape feature was composed of geometric features and improved moment invariant features, and the texture feature was composed of gray level co-occurrence matrix and local binary patterns (LBP) feature. The weighting matrix of color, shape and texture feature would be the input parameter after unitary processing. A three-step method for model updating consisting of model structure tuning, model parameter updating and model validation was presented in this article. Firstly, the deep belief networks (DBNs) of double hidden layers and single hidden layer were established. Secondly, the influence of the 3 types of constant, rising and descending nodes of double hidden layers in DBN was analyzed. The experimental result showed that the descending nodes of double hidden layers in DBN could learn the distributed characteristics of the original characteristic data better than the other node types of double hidden layers. Finally, the testing optimization parameters of double hidden layers and single hidden layer were obtained by experiment. The recognition rate of double hidden layers of DBN was 83.55% when the number of nodes stood at 101-210-55-6, and the recognition rate of single hidden layer of DBN was 91.13% when the number of nodes stood at 101-200-6. The DBN structure of single hidden layer was better able to excavate the distribution rule of weed features than DBN with double hidden layer. The single color, shape, texture and fusion feature were used to construct 3 types of weed classification models, which were support vector machine (SVM), BP (back propagation) neural network and DBN. In the experiment, the recogniton rate of DBN model with single color and shape feature was lower than that of the SVM and BP neural network model. The dimensions of color and shape feature were relatively small, which could not reflect the advantage of characteristic representation with DBN. On the other hand, the recognition of DBN model with single texture and fusion feature was more accurate than that of the SVM and BP neural network model, and the recognition rate of DBN model reached 86.58% and 91.13% with single texture and fusion feature, respectively. The results demonstrate that the method put forward in the paper can improve the classification accuracy of weeds with the complex background and variable illumination in paddy fields.