Study On Infield Weed Detection Using Machine Vision

The study on infield weed detection using machine vision is significant to reduce herbicide dosage and environmental pollution by applying the variable-rate spraying. However, as a new field of application, the technique of weed detection based on machine vision still has a long way to go in our country. Reviewing of relative research at home and abroad, a new method was studied to detect infield weed according to the information of position, shape and spectral feature of crop and weed and the system of infield weed detection based on machine vision was developed. The contents of the study could be briefly summarized as follows:1. The parallel and fast algorithms about capturing and processing dynamic images of infield weed detection were studied under the lighting, indoor and dynamic conditions. These algorithms provided the basis for the further research of real-time weed detection in preprocessing, segmenting and post-processing.2.The classified statistic analysis method was introduced into segmenting color images, which could be applied to segment color images that were composed of complicated scene.3.Because of the severe occluding of drilling crop leaves, it was difficult to extract shape and texture feature, so a new approach to detect the between-rows weed was discussed on the basis of the position feature, where crop was regularly sown as a constant row space and most weed were distributed on the bare-soil between crop rows during 3 leaves to 5 leaves seedling stage.4.The improved pixel lateral histogram algorithm was used to extract the centre of crop row. The infield adaptability of the pixel lateral histogram algorithm was enhanced by subsection computation of pixel lateral histogram.5.The seed fill algorithm in graphics was introduced into the between-rows weed detection to fill the areas connected with the centre of the crop row. In order to gain the faster processing speed, an improved scan-line seed fill algorithm was developed successfully.6.The algorithms of morphological operators and label watershed segmentation based on the mathematical morphology were put forward to resolve the problem of the light occluding of dibbling crop leaves. Compared with the traditional algorithm of watershed segmentation, the algorithm of label watershed segmentation was improved in the segmentation effect and executed time.7.In terms of the leaves' shape difference of corn seedling and weed seedling in the field, the approach that used shape feature factors of area and compactness to classify corn and weed was studied.8.Because the detection method using multi-spectral feature to identify weed is superior to other methods in real-time detection, the spectral characteristics of wheat and four kind of weeds from 666nm to 1176nm were studied. The results showed that it was feasible to use the multi-spectral feature of plants to detect infield weed.9. A test system of infield weed detection based on machine vision was designed and developed. The software system consisted of file management, system initialization, image preprocessing, background segmentation and weed detection.