CAJ | 학술논문

更加细致的体现病害外部形态特征和较为完好的保留病害区域颜色纹理信息，是玉米等作物病害分割的关键性研究问题之一。该文提出一种基于改进人工蜂群算法的脉冲耦合神经网络图像分割算法，该算法以最大香农熵和最小交叉熵加权线性组合作为蜂群算法收益度评价函数，通过引入尺度因子调整引领蜂和跟随蜂的解搜索策略，改进后人工蜂群算法与脉冲耦合神经网络相结合，实现网络参数的自动优化调节。在RGB色彩子空间上将该算法用于一组玉米常见病害彩色图像分割，并借鉴利用彩色图像合并策略得到最终病害分割结果。试验表明，该文算法较为细致的体现病害外部形态特征，较为完好的保留了颜色纹理信息；利用分割区域色度误分度 V(I)值作为评判标准，该文算法V(I)幅值顺次降低2.03%、7.05%、10.15%和11.2%，综合降低了7.32%也优于对比算法。因此，该文算法为病害彩色图像分割提供了一种较为有效的方法。
경가세치적체현병해외부형태특정화교위완호적보류병해구역안색문리신식，시옥미등작물병해분할적관건성연구문제지일。해문제출일충기우개진인공봉군산법적맥충우합신경망락도상분할산법，해산법이최대향농적화최소교차적가권선성조합작위봉군산법수익도평개함수，통과인입척도인자조정인령봉화근수봉적해수색책략，개진후인공봉군산법여맥충우합신경망락상결합，실현망락삼수적자동우화조절。재RGB색채자공간상장해산법용우일조옥미상견병해채색도상분할，병차감이용채색도상합병책략득도최종병해분할결과。시험표명，해문산법교위세치적체현병해외부형태특정，교위완호적보류료안색문리신식；이용분할구역색도오분도 V(I)치작위평판표준，해문산법V(I)폭치순차강저2.03%、7.05%、10.15%화11.2%，종합강저료7.32%야우우대비산법。인차，해문산법위병해채색도상분할제공료일충교위유효적방법。
The image segmentation of crop diseases is one of the critical technical aspects of digital image processing technology for disease recognition. However, because of background information complexity of crop disease images, boundary area vagueness and noise effect of light and vein texture, there is no robust easy and practical method. At the same time, the color texture feature is one of the important criteria for identifying diseases, but there are serious influences on feature extraction and disease recognition because of the color texture information ignorance of most of the methods at present. The main contribution of this paper is that the segmentation appearance is more subtle and the color texture information is better when kept in the target area of crop diseases based on the proposed method——a pulse coupled neural network based on a modified artificial bee algorithm (MABC-PCNN). The basic idea of the color disease image segmentation is that the method of MABC-PCNN was used to segment the disease regions in RGB subspaces, then the results in three subspaces were merged in reference to a selective large probability merge strategy, and finally the final merger result was obtained. The concrete realization is as follows. Firstly, a method of MABC-OCNN was proposed in this paper, and in this method the parameters of PCNN (βis the linking strength, Vθis an amplitude coefficient and aθis a an incentive pulse attenuation coefficient, Vθand aθ set the operation of neuromine) were automatically optimized through an improved ABC (MABC). In more detail, the above mentioned coefficient was described as the components of the feasible solution corresponding to the nectar source. By introducing scale adjustment factor , the solution search strategy of leader and follower had been adjusted, then through the evaluation principle of a weighted linear combination of maximum Shannon entropy and minimum cross-entropy, the results of segmentation with PCNN were evaluated and in the iteration of MABC, the optimal solution was set as the coefficients of PCNN. Secondly, in the iteration of the method, we got the optimal parameters of PCNN, and meanwhile we got the segmentation results in RGB subspaces. According to the selective large probability merge strategy, the results were merged and the final result of the color disease image segmentation was gotten. Further details are as follows that the pixel value variances of segmentation results in RGB subspaces were calculated, and then with the above variances the contributions of the pixel values were calculated. Finally, a mask template was obtained with the components of pixel value contribution. By masking operation with the mask template and an original color disease image, the final segmentation result was gotten. In a group of color maize disease images, the experimental results show that no matter whether subjective evaluation or objective evaluation was compared with V values, the segmentation appearance is more subtle and the color texture information is better remained in the target area of maize diseases based on the proposed method in contrast to GA-PCNN in Ref[18]. However, because of stochastic optimal control parameters with a swarm intelligent algorithm, the algorithm in this paper is of relatively high time complexity. Along with the continuous improvement of hardware performance, this problem will be solved.