| 摘要: |
| 单位面积的穗数是估算小麦产量的重要指标，针对传统麦穗计数方法效率低、主观性高等问题，将基于深度残差网络的密度回归模型引入麦穗的计数领域，建立原始图片与密度图的对应关系，以密度图像素值总和确定图像中麦穗数量。对ResNet34网络进行改进，提出了ResNet-16模型，实现端对端的麦穗计数。针对ResNet34网络复杂度高的特点，ResNet-16增加了残差块的宽度，减少了ResNet34网络的深度；为了避免真值密度图的精度误差以及梯度下降过快，引入了矫正因子δ和膨胀因子K。结果表明:改进后的ResNet-16模型能够取得更好的预测精度，平均绝对误差为2.50，均方根误差为3.27，相关系数R2为0.973，计数准确率达到94%，较MCNN计数模型精度提高了6%，可以实现高效、快速的麦穗计数。利用基于深度残差网络的回归计数模型为麦穗计数提供了一种新的计数方式。 |
| 关键词: 麦穗 计数 密度回归 残差网络 |
| DOI： |
| 投稿时间：2020-09-21 |
| 基金项目:中国农业科学院科技创新工程项目(CAAS-ASTIP-2020-AII)；中央级公益性科研院所基本科研业务费专项(Y2020XK07、JBYW-AII-2020-10) |
|
| Research on wheat ear regression counting based on deep residual network |
|
LIU Hang1,LIU Tao2,LI Shijuan1,LI Luhua1,LV Chunyang1,LIU Shengping1*
|
| (1.Institute of Agricultural Information, Chinese Academy of Agricultural Sciences, Beijing 100081, China;2.College of Agriculture, Yangzhou University, Yangzhou 225009, China) |
| Abstract: |
| The number of wheat ears per unit area is an important indicator for estimating wheat yield. In view of the low efficiency and high subjectivity of traditional wheat ear counting methods, a density regression model based on the deep residual network is introduced into the field of wheat ear counting. It establishes the corresponding relation-ship between the original image and the density map, and determines the number of wheat ears in the image by the sum of the pixel values of the density map. The ResNet-16 model is proposed on the basis of improving the ResNet34 model. In view of the high complexity of the ResNet34 network, ResNet-16 model is used to increases the width of the residual block and reduces the depth of the ResNet34 network; In order to avoid distortion of true value density map and too fast gradient descent, correction factor δ and the expansion factor K are introduced. The results show that: The improved ResNet-16 model can obtain better prediction accuracy. Its mean absolute error is 2. 50, root mean square error is 3. 27, correlation coefficient R2 is 0. 973, and the counting accuracy rate reaches 94%. Compared with MCNN model, the accuracy of ResNet-16 is improved by 6%, which can achieve high efficiency and fast ear counting. The regression counting model based on deep residual network provides a new counting method for wheat ear counting. |
| Key words: wheat ear counting density map regression residual network |
