利用融合高度与单目图像特征的支持向量机模型识别杂草

除草是保证农作物高产的必要工作。针对机械化除草和智能喷药中存在的杂草识别问题,以2~5叶苗期玉米及杂草为研究对象,进行了融合高度特征与单目图像特征的杂草识别方法研究。首先从单目图像中提取16个形态特征和2个纹理特征;然后基于双目图像,提出了针对植株的高度特征提取方法,所得高度特征与实际测量值间误差在±12 mm以内;利用max-min ant system算法对形态特征进行优化选择,将形态特征减少到6个,有效减少数据量62.5%,并与纹理和高度特征进行融合;将2~5叶玉米幼苗的可除草期划分为3个阶段,分别构建融合高度特征与单目图像特征的SVM识别模型,并与相应不含高度特征模型进行对比。经测试,3个阶段模型的识别准确率分别为96.67%,100%,98.33%;平均识别准确率达98.33%。不含高度特征模型的识别准确率分别为93.33%,91.67%,95%;平均识别准确率为93.33%。结果表明,融合高度特征与单目图像特征的SVM识别模型优于不含高度特征模型,平均识别准确率提高了5百分点。该方法实现了高准确率的杂草识别,研究结果为农业精确除草的发展提供参考。

Weed recognition using SVM model with fusion height and monocular image features

The technology of weed recognition based on machine vision becomes the research focus of precision agriculture.In order to realize the precise weeding technology, it is required to recognize weeds and crops rapidly and precisely.In this research, the high accurate recognition method of weed was studied.Maize seedlings of 2 to 5 leaves stage and weed during same stage were used as research object and method of accurate recognition of weed based on SVM recognition model that fusion height feature and image features was studied.We found that maize seedlings were generally higher than the weeds during the same period.This fact could allow us to have a more accurate recognition evidence for the SVM recognition model.In this paper, we conducted accuracy analysis.Binocular vision system was built and calibrated, an image acquisition system was designed and binocular images of crops and weeds were grabbed.The calibration results of binocular camera were evaluated using re-projection error.The average re-projection error was 0.08 pixels, and no more than 0.1 pixels.This demonstrated that the binocular image acquisition system had a high accuracy of the calibration processing. We also pretreated the collected images with methods of Excess-green feature, Improved Otsu, Area filtering, Canny edge detection.Sixteen morphological features and two texture features in monocular image based on image pretreatment were obtained.Then we proposed the height feature extraction method of plant.Based on binocular images, the height feature of plant can be extracted by this method quickly.Firstly, with height feature extraction method, the binocular gray image was masked with binocular binary image.With this processing on binocular image, the background information irrelevant to the recognition target was removed and the processing speed of the algorithm was improved.Then the speed up robust feature algorithm was used to detect the feature points in the masked binocular gray image.The method had good robustness.Matching feature points were detected between left and right images by the sum of absolute differences method.Those matching feature points that were not suitable for constraint of epipolar line were removed.Projective transformation matrix was worked out using basic matrix and coordinate of matching feature points.Moreover, the binocular images were corrected based on projective transformation matrix.With these processing, the left image and right image were corrected to the same horizontal epipolar line, the bad effect of mechanical vibration on image quality was reduced, and the robustness of the method in this paper was significantly improved.Next, the accurate disparity map was obtained by combining global error energy minimization algorithm with the average error threshold filter.Matching block in the size of 5 × 5 was selected in this processing.In this condition, processing speed of the method was maximized, and calculation precision was satisfied. Finally, according to the geometrical principle of binocular vision, the depth information was calculated based on the disparity map.The maximum depth information of the target area was referred to the height feature of plant.The error between actual height and calculation was less than ±12mm.After that, we divided weeding period into three stages and built SVM recognition model with fusion height feature and image features for each stage.Based on the max-min ant system algorithm, we selected the optimal morphological features of each model.As a result, the morphological features were reduced from 16 to 6, feature data was reduced by 62.5%.By using three kinds of algorithms: genetic algorithm, k-fold cross validation, and particle swarm optimization, the two core parameters of each SVM recognition model were optimized, and the optimal parameters were selected by comparing effects of three algorithms.The optimal parameters c and g contained by this way can effectively improve the recognition ability of SVM model, and avoid the over-learning and under-learning state of SVM model.The test set was used to test the recognition rate of SVM model of each stage, and the contrast experiment of the SVM model that did not fusion height feature and image features was carried out.After all those processes the weed recognition system was established.It was necessary to segment the overlapping images of maize and weed.Watershed algorithm based on distance transform was used to solve this problem.Firstly, the segmentation line was determined by using watershed algorithm based on distance transformation.Then the overlapping image was segmented by morphology processing for maintain the original shape.Finally, the binary image after segmentation was adopted to mask the gray image and the segmentation of weeds and crops was realized.This method can effectively segment maize and weeds in the seedling stage.The results of this research showed that the recognition rate of the SVM model based on fusion height feature was 96.67%, 100%, 98.33%, with the average recognition rate of 98.33%.The recognition rate of the SVM model based on no fusion height feature was 93.33%, 91.67%, 95%, with the average recognition rate of 93.33%.The data showed that the SVM recognition model based on fusion height feature was better than the model without fusion height feature, and the average recognition rate was improved by 5%.Therefore, weed recognition method based on fusion height feature and SVM model can effectively improve the recognition rate and achieve high accuracy of weed recognition.The research results in this paper will provide reference for the development of precision weeding.