Wheat Ear Identification Technology Based On Insufficiency Information

At Australia phenotype plants and bio-informatics and phenotype Research Center University of South Australia, wheat varieties were breeded to possess fine environmental adaptability by mean of plant phenotypics based on computer vision technology. After years of long-term observation, multi-spectral image of wheat growthing have accumulated rich, they plan to ananlysis the ear of wheat phenotypic in order to achieve the purpose of cultivating high-yielding wheat varieties.In the image analysis of the phenotype of wheat, wheat is the subject, its information by means of the imaging device were fully expressed. But wheat ear as part of the wheat does not belong to the body imaging, serious expression of their information is not sufficient, and therefore belong to incomplete information.In this paper, the object recognition technology based on the computer vision were analysised, the results showed that wheat ear couldn’t be identified bythe currently object recognition technology. But the approach that Wheat ear texture analysis can be used. In spatial and frequency domain, the identification of wheat ear were studied in this paper.In the spatial domain analysis, based on an analysis of wheat multispectral image, the implement method of Laws texture energy was proposed to detec the wheat ear. By programming, data processing, the model using neural network and support vector machine were established. After comparison and verification of the model, the studies show incomplete information in the background, it was impossible to achieve to identify wheat ear by single discriminant model, and the model structure must be hierarchical:Wheat ear area determination model based on weak classifiers and Wheat ear recognition model based on strong classifier. Hierarchical model structure reached 88.75% accuracy rate in all test images.According to the study of computer visiual saliency, the paper transformed from the spatial domain to meet the human visual characteristics Gabor frequency domain, spatial domain identification convert to identify the problem in the frequency domain. Different from the usual focus on saliency research on the low-frequency part of the whole image, the paper will examine in a high frequency, the idea put forward in the frequency domain using the "frequency domain residual", "partial" change the traditional frequency domain processing method to achieve the wheat phenotypes image under complex conditions in the Wheat ear field identified by Gabor local significance testing. Through the design process, validation analysis, and ultimately determine the parameters and the frequency range of the frequency domain Gabor, gives significant test parameters.After the image of the test results show that the local domain Gabor significant way to achieve a good Wheat ear area determining requirements under complex conditions. Use this method as a first-level hierarchical model can greatly enhance the performance of discriminating Wheat ear region.After the first stage of the regional classification model was sured to use the Gabor domain of local saliency, the paper has conducted research for a strong second-stage classification model. On the basis of analysis of the current regional characterization methods, determined using local binary feature (LBP) as a regional feature. After some research on the traditional machine learning methods chosen to support vector machine (SVM) classifier as a strong model. Through the design and experimental procedures, and ultimately determine the parameters of LBP and SVM parameters.Hierarchical model was used to check all the test images, eventually reaching 91.37 percent correct. Also pointed out that due to incomplete information, the current wheat phenotypes images only when the Wheat ear to 6.5mm long time in order to be detected.Relative to shallow traditional learning methods, this paper proposes a method for in-depth study Identify Wheat ear incomplete Environment in. Through the study of the depth of learning, with the depth of the model as a strong hierarchical model of classification model. The accuracy of the model reached 98% and the processing time of 0.06ms. It shows the depth of learning strong analytical skills.