基于自适应三分图的玉米穗丝精细化提取

玉米穗丝是玉米的受粉器官，它的生长发育状况将直接影响玉米产量。为了实现穗丝性状的自动监测与评价，提出一种基于自适应三分图的玉米穗丝精细化提取方法。首先基于直方图对比度的显著性检测算法预估目标显著性，形成显著图，再利用自适应阈值分割对显著图初步预分割；然后寻找最大连通域获得单一目标二值图像，根据穗丝特点确定形态学处理的最佳卷积核并生成包含目标、背景和不确定区域的自适应三分图，以Laplace矩阵优化closed-form matting算法中目标函数的分块矩阵；最后根据玉米穗丝特点，结合三分图结果和优化过的算法对玉米穗丝进行软分割实现精细化提取。实验结果表明：所提的研究方法像素精度值为97.96%,综合评价指标值为94.16%;其中像素精度值相较于基于点运算的OTSU算法、Grabcut算法和基于神经网络的深度学习算法SeFormer和DeeplabV3,分别提高了10.43、2.35、1.67和1.91百分点，综合评价指标值分别提高了23.96、5.59、4.28和5.13百分点。此方法能有效弥补硬分割算法对玉米穗丝这种特殊对象的分割缺陷，提取的目标区域接近真实区域，准确度高，从而为智...

Refined Segmentation of Corn Silk Based on Adaptive Trimap

The corn silk is the pollinating organ of maize, and its growth and development are closely related to the yield. In order to realize the automatic monitoring and evaluation of corn silk characters, proposing a refined extraction method of corn silk based on adaptive trimap. First, the significance detection algorithm based on histogram contrast is used to estimate the significance of the target and form a saliency image. Then, the saliency image is preliminarily pre segmented by using adaptive threshold segmentation, and finding the largest general domain to obtain a single target binary image. According to the characteristics of the silk, determining the best convolution kernel for morphological processing, and generating a trimap containing the background and uncertain regions. Laplace matrix is used to optimize the block matrix of the objective function in the closed form matching algorithm. Finally, according to the characteristics of the corn silk, the soft segmentation of the corn silk is carried out by combining the results of the trimap and the optimized algorithm to achieve refined extraction. The pixel accuracy value of this research method is 97.96%, and the comprehensive evaluation index F-measure value is 94.16%. Compared with OTSU algorithm and Grabcut algorithm based on point operation, and deep learning algorithms SeFormer and DeeplabV3 based on neural network, the accuracy is increased by 10.43, 2.35, 1.67 and 1.91 percentage points respectively, and the F-measure value is increased by 23.96, 5.59, 4.28 and 5.13 percentage points respectively. This method can effectively remedy the defect of hard segmentation algorithm in the segmentation of special objects such as corn ear silk. The extracted target area is close to its real area with high accuracy, thus providing technical support for intelligent monitoring of corn silk growth.