Stalk tissue anatomy characteristic parameter high-throughput extraction method based on deep learning

Abstract

The invention provides a stalk tissue anatomy characteristic parameter high-throughput extraction method based on deep learning. The method comprises the following steps: S1, drawing a contour and a label category for a digital slice image; S2, using ResNet as a feature extractor, mapping the original image to a high-dimensional feature space, and then outputting feature maps on multiple levels; S3, extracting a high-order feature map and a low-order feature map by using a feature intensifier of an encoder-decoder structure, and fusing the high-order feature map and the low-order feature map to obtain a strongest feature; S4, in order to identify the instances in the same semantic category, indicating different instances by using a branch detection method, wherein the branch detection method comprises the following steps: adopting a category branch, a width-height branch and a regression branch to generate detection output; S5, after detection output, predicting the contour of the vascular bundle by using an instance mask, and representing three functional areas of the cross section of the corn stalk by using a semantic mask. According to the method, phenotypic parameters of the cross section slice image of the corn stalk can be analyzed and detected highly accurately and robustly.

Description

technical field [0001] The invention relates to the field of deep learning, in particular to a high-throughput extraction method for anatomical characteristic parameters of stem tissue based on deep learning. Background technique [0002] Microscopic phenotypic features in crop stems, especially vascular bundles and functional domains, are closely related to nutrient transport and lodging resistance in crops. Quantitative analysis of anatomical features using crop cross-sectional slice images is a great challenge for scientists in plant science. Therefore, there is an urgent need to develop an accurate and high-throughput phenotyping method. Traditional image processing methods are not suitable for processing uncontrollable but non-negligible variable factors in maize stalk section images, and high-throughput methods based on micro-CT technology are expensive, which limits their application. Deep learning methods simulate the human eye to observe, learn features of objects...

AI Technical Summary

Problems solved by technology

However, there are few reports of deep learning-based m

Images