Method for quickly counting rice ear number of field rice by using image pyramid and Fast-RCNN

Abstract

The invention discloses a field rice spike counting algorithm HW-Faster-RCNN based on an image pyramid. The field rice spike counting algorithm HW-Faster-RCNN realizes automatic counting of field riceimages with relatively small rice spike sizes and more than 50 rice spikes. Firstly, the size of a receptive field is analyzed, an appropriate feature learning network is selected, and an appropriateinput image size is calculated according to the rice spike size; secondly, an original image is cut based on the input image size range, and an image pyramid is designed and constructed to increase the relative size of rice ears; then, a mixed window data set is constructed based on the image pyramid, and a rice ear counting model is trained by using an HW-Faster-RCNN network; and finally, the to-be-tested image is segmented, the sub-images are respectively counted by using a rice spike counting model, and a damaged spike fusion algorithm is designed to remove repeated counting. The result shows that the counting accuracy of the small-size rice ears is remarkably improved by selecting the feature learning network and designing the image pyramid according to the ear sizes; the average counting accuracy and the error detection rate of the HW-Faster-RCNN are 87.23% and 4.60% respectively, and the HW-Faster-RCNN can be practically applied to automatic counting of rice ears in a complex field scene.

Description

technical field [0001] The invention belongs to the technical field of plant phenotype detection, and is an intersecting field of an image pyramid algorithm, a target detection algorithm based on deep learning and automatic detection of quantitative traits of rice ears. This paper proposes a comprehensive method for optimal feature learning network selection, image pyramid design and construction, and automatic counting of a single image with more than 50 rice spikes that can be used in the mature stage of rice grown naturally in the field. Background technique [0002] Rice is one of the main food crops in Asia. Accurate output estimation can provide a basis for scientific production decision-making. The yield of cereal crops such as rice is mainly determined by three agronomic indicators, the number of spikes per unit area (panicle density), the number of grains per spike, and the thousand-grain weight. Previous studies have shown that the number of grains per unit area ...

AI Technical Summary

Problems solved by technology

, the growth environment of rice in the field is complex, and there are a large number of rice ears and leaves per unit area under high planting density, and the occlusion and bonding between rice ears and leaves are serious; the size of rice ears is relatively small; different from wheat, mature The rice spikes in the early stage are drooping due to their own weight, and the uneven illumination leads to huge color differences between different rice spikes; these all make it difficult to automatically count rice spikes in the field based on images

In addition, the existing color or texture-based segmentation methods and candidate region-based classification methods mainly focus on scenarios where the number of spikes in a single image is less than 30, when the number of rice spikes to be counted and irrelevant leaves, weeds, etc. increase Time accuracy is generally not high

Images