吴一全, 殷骏, 戴一冕, 袁永明. 基于蜂群优化多核支持向量机的淡水鱼种类识别[J]. 农业工程学报, 2014, 30(16): 312-319. DOI: doi:10.3969/j.issn.1002-6819.2014.16.040
    引用本文: 吴一全, 殷骏, 戴一冕, 袁永明. 基于蜂群优化多核支持向量机的淡水鱼种类识别[J]. 农业工程学报, 2014, 30(16): 312-319. DOI: doi:10.3969/j.issn.1002-6819.2014.16.040
    Wu Yiquan, Yin Jun, Dai Yimian, Yuan Yongming. Identification method of freshwater fish species using multi-kernel support vector machine with bee colony optimization[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2014, 30(16): 312-319. DOI: doi:10.3969/j.issn.1002-6819.2014.16.040
    Citation: Wu Yiquan, Yin Jun, Dai Yimian, Yuan Yongming. Identification method of freshwater fish species using multi-kernel support vector machine with bee colony optimization[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2014, 30(16): 312-319. DOI: doi:10.3969/j.issn.1002-6819.2014.16.040

    基于蜂群优化多核支持向量机的淡水鱼种类识别

    Identification method of freshwater fish species using multi-kernel support vector machine with bee colony optimization

    • 摘要: 为了准确地进行淡水鱼种类自动识别,利用计算机视觉技术,提出了一种基于Krawtchouk矩、灰度共生矩阵、蜂群优化多核最小二乘支持向量机(least squares support vector machine,LS-SVM)的识别方法。首先获取淡水鱼样本的灰度图像,计算淡水鱼鱼体的长宽比、鱼头鱼尾的Krawtchouk矩不变量形状特征,求得鱼身的灰度共生矩阵纹理特征,将上述形状与纹理特征组合成高维特征向量,并输入到多核LS-SVM,通过人工蜂群(artificial bee colony,ABC)算法对多核LS-SVM中的待定参数进行寻优,ABC算法中的适应度函数为测试样本的识别精度;最后输出识别精度达到最高时的最优参数。利用该方法对鳊鱼、鳙鱼、鲫鱼、草鱼、青鱼5种淡水鱼进行了分类识别,对鳊鱼、鳙鱼、鲫鱼、草鱼4种鱼识别时,各类鱼的识别精度均达到95.83%以上,对鳊鱼、鳙鱼、鲫鱼、青鱼4种鱼识别时,各类鱼的识别精度均达到91.67%以上,对鳊鱼、鳙鱼、鲫鱼、草鱼和青鱼 5种鱼识别时,各类鱼的识别精度均达到83.33%以上;与近年来提出的淡水鱼识别方法、BP(back propagation)神经网络方法、单核LS-SVM方法相比,该方法的识别精度更高,从而可快速准确地识别淡水鱼的种类,提高水产养殖的自动化水平。

       

      Abstract: Abstract: Currently, the identification of freshwater fish species is conducted by people. Computer vision techniques can solve problems existing in identification by hand, such as a large amount of work and low precision. Firstly, freshwater fish are converted into an image signal. Then images are transmitted to the computer processing system to be analyzed and processed. The features of freshwater fish images are extracted, and the movements of the equipment are controlled according to the features. As a result, the production efficiency and production quality are improved by this substitute for the manual work. Thus, in order to identify freshwater fish species automatically and accurately, computer vision techniques are adopted and an identification method based on Krawtchouk moments, a gray level co-occurrence matrix, and a multi-kernel least squares support vector machine (LS-SVM) with bee colony optimization is proposed. Firstly, gray level images of freshwater fish samples are obtained by camera, image acquisition card, and PC. Secondly, shape features such as the length to width ratio of freshwater fish and Krawtchouk moment invariants are calculated. The length to width ratio can be obtained by looking for the minimum circumscribed rectangle of fish. The fish head, fish body, and fish tail are obtained based on the minimum circumscribed rectangle. Then the Krawtchouk moment invariants of the fish head and those of the fish tail can be obtained. Thirdly, four characteristic parameters such as angular second moment, entropy, deficit moment, and variance based on a gray level co-occurrence matrix of the fish body are calculated, which are used as texture features. Finally, the above shape features and texture features are combined into a high dimensional feature vector. The feature vector is input into multi-kernel LS-SVM. In a multi-kernel LS-SVM, a polynomial kernel function and radial basis function kernel are used, which can make up for limitations of a single kernel function. The artificial bee colony algorithm is used to optimize the undetermined parameters in a multi-kernel LS-SVM. The identification accuracy of the test sample serves as a fitness function. The optimum parameters are output once the identification accuracy attains its maximum. Five freshwater fish species such as bream, bighead, crucian carp, grass carp, and black carp are identified by using the proposed method. A total of 190 fish were used as the research object. Among them, 70 fish were used as training samples, and the remaining 120 fish were used as testing samples. The experimental results showed that when bream, bighead, crucian carp, and grass carp are identified, the identification accuracy of them can all reach more than 95.83%. When bream, bighead, crucian carp, and black carp are identified, the identification accuracy of them can all reach more than 91.67%. When bream, bighead, crucian carp, grass carp, and black carp are identified, the identification accuracy of them can all reach more than 83.33%. In comparison with the freshwater fish identification method proposed recently, the back propagation (BP) neural network method, and the single kernel LS-SVM method, the proposed method has the higher identification accuracy. It indicates that the above shape features and texture features extracted can reflect the information of freshwater fish more comprehensively and effectively. Furthermore, the performance of a multi-kernel LS-SVM is much better than that of a single kernel LS-SVM. The proposed method is expected to find wide applications in a freshwater fish processing system.

       

    /

    返回文章
    返回