Research On Weed Seeds Recognition Using Manifold Learning

Agriculture is the crucial industry of national economy, and it is the foundation for economic construction and development. Therefore, improving the quality and output of agricultural products can contribute significantly to the economic development. Weeds can seriously affect the growth and development of crops, which will lead to low crop yields and inferior quality of crop. So the recognition of weed seeds becomes vitally important. With the development of information technology, application of machine learning and image recognition to agriculture becomes more and more popular, especially in the recognition and quality evaluation of maize and wheat seeds. However, the application to weed seed recognition is still new areas, and relative research is very little. Based on summing up the domestic and oversea relative researches, this paper propose a series of novel techniques for recognition of weed seeds, known as manifold learning algorithms and principal component analysis. Manifold learning algorithms and principal component analysis are used to extract features, which are used by some classifiers to recognize weed seeds.The main contributions are as follows:(1) Acquisition and preprocessing of Weed seeds data setThis study uses a weed seeds database as the experimental data set. The database has 50 categories weed seeds with 20 samples for each category. All of the weed seeds images should be preprocessed before being used, such as size transformation, unifying background color and adjusting the location of the seed images.(2) Researches on feature extraction algorithm of weed seedsThis study does researches on traditional feature extraction: principal component analysis and the current hot manifold learning algorithms: locally linear embedding and isometric mapping. These methods are applied to feature extraction of weed seeds to get d dimensional feature, and then these features are applied to three kinds of classifiers to get recognition results, at last the experimental results will be analyzed.(3) Researches on parameters setting of manifold learning algorithm.This study does researches on parameters setting of two manifold learning algorithms: LLE and ISOMAP. From experimental point of view, this study analyzes the recognition rates under d dimention using different k, and ultimately determines k and d. (4) Recognition of weed seedsThe experimental data set consists of 50 weed seeds classes with a total of 1000 color images. Four algorithms are used to extract features of the data set, and then these features are applied to three kinds of classifiers to get recognition rates. Experimental results demonstrate that the total recognition rate reaches 97%, and meets the design requirements, so manifold learning algorithms and principal component analysis can be used to extract features of weed seeds.