Research On Identification Of Nightshade Crops And Weeds Based On Machine Vision

Based on computer machine vision system, the main purpose of this study was to identify weed from nightshade by using digital image processing and pattern recognition. Specific study was carried on under normal sunlight condition by using digital camera photo in order to identify weed from nightshade. The main contents of the study were as follows:(1) According the need of color features by successive process, the original images were preprocessed by using color image filtration. A comparison was drew between scalar filtration and vector filtration. The results showed that color image scalar filtration excel vector filtration considering the effect and speed of image preprocessing.(2) In the image segmentation part, three methods (including histogram, maximum variance threshold, optimal threshold) were applied to segment the object and background with the idea of Excess green (GE.Meyer). The results showed that optimal threshold method has the best effect.(3) The problem of light condition in segmentation was solved by using color component normalization. For the problem of leave coverage in edge detect, the author used the idea of chromatism (1v.Mingzhong), which use the color gradient instead of gray-scale gradient in edge detect, to improve on Sobel edge detector. The correctness is over 70%. And mathematical morphological method was used to deal with the noises in binary image.(4) In allusion to the features of nightshade and weed, the author studied four nondimensional shape features, including aspect, first invariant moment, elongatedness and compactness, and diagrammatized, found that the optimal time for identification was 27^th after burgeoning. In the part of texture feature, the author tested the method of statistical gray-scale difference (using CON and ASM). The diagram showed that object and weed can be identified in the rough.(5) A back-propagation neuron network was designed for weed identification, with the modifying of activation function and performance function in connotativelayer. According to the experiments, the optimal network structure was 6-12-2 with the training goal 0.03, training speed 0.01. Its identification rates of brinjaul, green pepper, stephanotis are respectively 96%, 94%, 100%, 96% and 94%.The aim of the research is to provide essential theoretical evidence and technological basis for the further development of herbicide use which is valuable in environment protection. The results of this paper attach great importance to reduce the gap between China and developed country in the area of agricultural information automation technology. It accelerates the application of computer image processing technology in agriculture engineering field.