Research Of Weed Recognition Method Based On Feature Optimization And Multi-feature Fusion

Crops suffering from the weeds in farmland is one of the basic questions impeding the growth of crops. In traditional agriculture, the main weeding method by spraying herbicide extensively without distinguishing crops and weeds would not only result in waste of herbicide and labor forces, but also cause environment pollution and health hazards. Precision farming, which is an integration of electronic, computer techniques, information techniques and intelligent mechanics, has been an inevitable trend of modern agriculture. The purpose of precision farming is to reduce pollution, protect ecology and promote the sustainable development of agriculture. One of the key technologies for precision farming is to identify the weeds from background and spray herbicide quantitatively and automatically based on machine vision and image processing. Identification is the primary technique and also the bottleneck for implementation of herbicide spot-spraying. In order to improve accuracy and efficiency of weed recognition, it is necessary to study new ideas, new methods of intelligent recognition patterns and new image processing hardware etc.. In this dissertation, study of weed identification in cotton field was carried on by using image processing, intelligent recognition, embedded platform and other technology. The main contents of the study could be briefly summarized as follows:1. Summarized the recent development of the studies at home and abroad on weed recognition based on color, shape, texture, and multi-features respectively by using computer vision technology. Analyzed and pointed out the problems of these studies and proposed a recognition method based on feature optimization and multi-feature fusion.2. Five dominant species of weeds including digitaria sanguinalis, eclipta prostrata, veronica persica, acalypha australis and portulaca oleracea in cotton fields of Jiangsu northern area were selected as research objects. First, applying a series of pre-processing such as image enhancement and filtering to acquired images. Then, dividing green plants from the background through automatic threshold method. Finally, the segmentation methods of overlapped leaves were specially studied: according to the difference of shape and degree of overlapping, these leaves were separated respectively through morphology operation or threshold segmentation based on distance transformation and watershed algorithm.3. On the basis of image processing, feature selection method of color, shape and texture of plant leaf were studied respectively.3 low-order color moments, which were used to describe the average color, color variance and color shifts, were extracted in HIS space as color features.17 shape features including geometric shape and Hu invariant moments were extracted from the contours of the single leaf.4 texture features such as energy, correlation, inertia and entropy were extracted based on gray level-gradient co-occurrence matrix which can obtain not only the internal information but also the edge information of the image.4. Based on wrapper-mode of feature selection, SVM algorithm was incorporated into ACO-based feature selection process so as to make full use of ability of SVM learning and advantages of swarm intelligence to solve difficult combinatorial optimization problems. Using the highest accuracy of classification by defining the evaluation function to guide the operation of feature selection to obtain optimum feature vector and hyper-parameters of SVM classifier simultaneously. The experimental results indicate that this method can chose those features which contribute most to classification and eliminate secondary or redundant features. As a result, the number of leaf's shape features reduced from 17 to 6 after optimizing, so the feature subset was effectively compacted and the input-dimension of machine-learning was significantly reduced. The accuracy this method achieved was about 90% by using optimized feature subset.5. A method combined SVM with D-S evidence theory based on multi-feature fusion was proposed. According to the low reliability and stability problem caused by some factors like light changes, leaves overlapping, error accumulation in image processing etc. of single feature-based method, recognition results based on color, shape and texture feature were regarded as 3 independent evidences to construct BPA function, and then, the final results were obtained by using the rules of combination and the thresholds of decision making. The experimental results have shown that the proposed method can obtain accuracy over 96% and it has good performance on accuracy and stability compared to the single type feature-based method in weed recognition.6. Weed recognition algorithms were realized on DSP platform through code transplant. First, time consuming of different algorithms was tested by using CCS testing tool, and then optimization of code and platform was made purposefully. Next, the performance under DSP platform was compared with PC platform, and the feasibility for implementation of time consuming algorithm by hardware was analyzed. Finally, the performance of media filter was validated based on FPGA and an embedded design scheme for weed recognition by a comprehensive development method was proposed.At the end of this dissertation, some conclusions are given and some further study areas are proposed.