基于多特征融合和SVM的玉米幼苗与杂草识别检测

Identification and Detection of Maize Seedlings and Weeds Based on Multi Feature Fusion and SVM

摘要: 杂草和作物的检测是使用喷洒除草剂机器人进行精确喷洒和田间农业机械精确施肥的关键步骤。基于使用颜色信息和连通区域分析的k-means聚类图像分割，提出了一种结合多特征融合和支持向量机(SVM)的方法来识别和检测玉米幼苗和杂草的位置，以减少杂草对玉米生长的危害，实现精确施肥，从而实现精确除草或施肥。首先建立玉米苗期杂草和玉米幼苗分类的图像数据集；其次提取玉米幼苗和杂草的许多不同特征，并通过主成分分析进行降维，包括定向梯度特征直方图、旋转不变局部二值模式(LBP)特征、HU不变矩特征、Gabor特征、灰度共生矩阵和灰度梯度共生矩阵；然后基于SVM进行分类器训练，得到玉米幼苗和杂草的识别模型，并对单个特征或不同融合策略特征的综合识别性能进行比较和分析，得出最佳特征融合策略；最后利用实际玉米苗期图像，测试所提出的杂草和玉米苗期检测方法的效果。实验结果表明，基于SVM分类器的旋转不变LBP特征和灰度梯度共生矩阵的融合特征组合获得了最高的分类精度，并准确地检测出各种杂草和玉米幼苗。它为喷洒除草剂的机器人提供杂草和作物位置的信息，以实现精确喷洒，或为精确施肥机提供信息，以进行精确施肥。

Abstract: The detection of weeds and crops is a key step in using herbicide spraying robots for precise spraying and precise fertilization of agricultural machinery in the field. Based on the use of color information and connected region analysis in k-means clustering image segmentation, a method combining multi feature fusion and support vector machine (SVM) is proposed to identify and detect the positions of corn seedlings and weeds, in order to reduce the harm of weeds to corn growth, achieve precise fertilization, and thus achieve precise weed control or fertilization. Firstly, establish an image dataset for the classification of weeds and maize seedlings during the maize seedling stage. Secondly, extract many different features of corn seedlings and weeds, and perform dimensionality reduction through principal component analysis, including directional gradient feature histogram, rotation invariant local binary pattern (LBP) feature, HU invariant moment feature, Gabor feature, gray level co-occurrence matrix, and gray level gradient co-occurrence matrix. Then, based on SVM, the classifier is trained to obtain recognition models for corn seedlings and weeds. And compare and analyze the comprehensive recognition performance of individual features or features with different fusion strategies to obtain the optimal feature fusion strategy. Finally, the effectiveness of the proposed weed and maize seedling detection method was tested using actual maize seedling images. The experimental results show that the fusion feature combination of rotation invariant LBP features and gray gradient co-occurrence matrix based on SVM classifier achieves the highest classification accuracy and accurately detects various weeds and corn seedlings. It provides information on weed and crop locations for robots spraying herbicides to achieve precise spraying, or information for precision fertilizing machines to carry out precise fertilization.