Field crop wide-view image generation method and system based on deep learning
By simultaneously acquiring data from multiple high-resolution RGB cameras and combining it with a cascaded network of SuperPoint and SuperGlue, the problems of limited field of view and severe distortion in traditional technologies are solved, generating high-precision field images that are adaptable to diverse crop scenarios and meet the needs of crop phenotypic analysis.
CN122244666APending Publication Date: 2026-06-19WUHAN GREENPHENO SCI & TECH CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN GREENPHENO SCI & TECH CO LTD
- Filing Date
- 2026-01-27
- Publication Date
- 2026-06-19
AI Technical Summary
⚠Technical Problem
Traditional technologies struggle to achieve high-precision field crop image acquisition, suffer from limited field of view, severe distortion, and insufficient generalization ability, failing to meet the needs of diverse crop scenarios.
⚗Method used
Multiple high-resolution RGB cameras are used to simultaneously acquire data and perform distortion correction. Combined with a SuperPoint and SuperGlue cascaded network, a targeted loss function is constructed through random homography transformation and perspective transformation to achieve image stitching and fusion.
🎯Benefits of technology
It generates wide-field, high-resolution field crop images, adaptable to various crops and different growth stages, improving the accuracy and efficiency of image analysis.
✦ Generated by Eureka AI based on patent content.
Abstract
This invention discloses a deep learning-based method for generating wide-field crop images, belonging to the field of smart agriculture technology. The method utilizes a mobile platform equipped with multiple high-resolution RGB cameras to simultaneously acquire top-view images of crops after calibration and distortion correction. A dataset is constructed and enhanced based on the acquired images, and a cascaded network of SuperPoint and SuperGlue is built. The former detects key points and generates descriptors, while the latter achieves feature matching through a graph neural network. Intrinsic points are selected using the RANSAC algorithm, the homography matrix is calculated, the images are aligned through perspective transformation, and multi-band fusion is used to eliminate seams, gradually stitching together to generate a wide-field image. This invention solves the problems of limited field of view, severe distortion, and insufficient generalization ability of traditional techniques, providing high-resolution and accurate crop images and reliable data support for crop phenotypic analysis.
Need to check novelty before this filing date? Find Prior Art