目的 基于模式识别的花卉种类识别方法在使用不同特征或分类器时识别准确率有较大差别。本文的研究目的在于实现花卉种类识别方法的快速构建及性能评估,减轻研究人员的编程工作量,提高效率。方法 根据使用模式识别技术进行花卉种类识别的一般步骤,应用插件技术将算法中的预处理、特征提取、分类器训练、分类器识别等步骤表示成不同种类的处理器,建立可扩展的系统平台,研究人员可以通过修改各步骤所使用的处理器来修改图像处理和识别算法,并在此基础上采用流式链接方法构建算法。结果 基于本文所提出的构建方法进行算法优化,并将其用于68种花卉的识别,准确率Top1为 91.26%,Top5为98.41%。结论 流式链接方法能够对识别方法进行快速装配,有利于快速评估不同特征和分类器在花卉种类识别中的性能,适于算法的研究和优化。本文所提出的基于工作流式链接方法以及插件技术的构建方法具有灵活易用的特点,所构建的算法具有良好的可扩展性。该方法还可以推广应用于其他基于数字图像的模式识别算法研究。

Objective The methods based on pattern recognition achieve quite different accuracy when using different features or classifiers. The purpose of research is to construct the flower category recognition method rapidly, measure its performance, reduce the development workload of researchers and improve efficiency. Method According to the normal procedure of flower category recognition using pattern recognition technology, the steps of the algorithm are programmed as different kinds of processors to build up an expandable system using the plug-in technology, and the researchers can change the image processing and recognition algorithm by choosing the corresponding processors. Based on this, the algorithm is constructed by linking the processor in the way of data flow and represented as a network. We use the XML description file to describe the dependencies of a plugin and object pool to provide the communication foundation between processors which is the base of the scalable construction methods. The processors are encapsulated into processor plugins and managed by the main framework. The processors are divided into three kinds of processors to input data, process data, and buffer data. The core processor plugin controls the execution order of processors and make sure the data inputted to the processor valid. The constructed processor network of flowers category recognition algorithm normally contains input-processor, feature extractors, buffer processor and classifier processor. The input-processor can generate one input data and data index each time. Feature extractors will be run for several times to process the input data sequence. In the procedure of classifier training or accuracy statistic, a feature vector buffer is needed. While the network that represents the algorithm is processing, the ports in processor can transport messages and data produced. Finally we compared and analyzed different features in color, shape and texture using K-Nearest Neighbor and support vector machine classifier respectively. We use HSV color space histogram to reduce the effect of light, Hu moment and edge curvature histogram to represent shape and polar gray-level co-occurrence matrix for texture. Then different accuracies are achieved by changed the parameter. We can use the constructed method to optimize the parameter of each flower feature extractor rapidly. After that, we combine the features which are extracted by flower extractor processors using the optimized parameter as a flower feature vector. At the same time, multithread technology is used to speed up the process of the algorithm constructed. Result We construct and optimize the algorithm using the method proposed in the paper. The recognition rate is 91.26% for first hypothesis and 98.41% for fifth hypothesis while using the optimized algorithm on a dataset of 68 flower species. We use the proposed method to construct a flower feature extraction algorithm on 853 training images and finish the whole procedure in 2125 s. Conclusion The connection method of workflow can construct an algorithm fast and benefit the rapid assessment of the performance of different features and classifiers in flowers category recognition. It is applicable to the research and optimization of algorithms. The proposed method based on workflow and plugin technology is easy to use and flexible and the algorithm constructed has good scalability. Furthermore, the system can be applied to the other researches using the methods of pattern recognition based on digital image.