Neural Network Image Generative Model for Depth Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image capture technologies fail to effectively represent the three-dimensional structure of real scenarios, as depth images do not accurately reflect the depth information of a scene, limiting their application in fields like 3D measurement and man-machine interaction.
Innovation Solution
A method and apparatus for generating an image generative model using a sample set comprising depth images and visible images, where the image resolutions are adjusted and input into a pre-established neural network model to generate high-resolution and high-quality depth images, with the model being trained to reach a preset optimization goal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If depth images are captured using conventional camera technology, then the three-dimensional structure information can be obtained, but the depth information accuracy is insufficient
Solution Approach 1:
The patent introduces a neural network model as an intermediary between the captured depth image and the final high-precision depth map. The model processes the low-quality depth image and visible image to generate a high-precision depth map, effectively mediating the transformation from inaccurate to accurate depth information
Solution Approach 2:
The patent performs preliminary actions by capturing both depth images and visible images simultaneously, and pre-processing them through the neural network model before final output. This preliminary processing ensures that the depth information is enhanced and corrected before being used for 3D reconstruction or measurement
2Manufacturing precision
If high-resolution depth images are processed directly, then the detail information is preserved, but the processing complexity and time consumption increase
Solution Approach 1:
The patent applies partial action by processing only the necessary features and regions of the depth image through the neural network model, rather than processing every pixel uniformly. This selective processing approach maintains high-resolution output while reducing overall processing time and computational resources required
Data Source
AI summary
Methods and apparatuses for generating an image generative model are disclosed. An embodiment comprises: acquiring a sample set, a sample comprising a first depth image, a second depth image and a visible image; and executing training based on the sample set: inputting the second depth image and the visible image of at least one sample in the sample set respectively into a pre-established initial neural network model to obtain a generated depth image corresponding to each of the at least one sample; calculating a similarity between the generated depth image corresponding to the each of the at least one sample and a corresponding first depth image; determining whether the initial neural network model reaches a preset optimization goal based on the calculation result; and using the initial neural network model as the trained image generative model, in response to determining the initial neural network model reaching the preset optimization goal.