Depth Estimation Model Training Using Autoencoder Image Restoration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current depth estimation methods in machine vision face reliability issues due to pixel differences between binocular images, leading to inaccurate predictions and reduced model reliability.
Innovation Solution
A method for training a depth estimation model that involves using an autoencoder and an image conversion module, along with camera calibration parameters, to process binocular images and minimize pixel value differences through a back-propagation algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If binocular images with pixel differences are input to a training model, then depth estimation can be performed, but the predicted value deviates and model reliability reduces
Solution Approach 1:
The patent applies preliminary action by performing image preprocessing operations before depth estimation. Specifically, it uses autoencoder-based image restoration to correct degradation in binocular images, and employs image fusion techniques to align and harmonize pixels from left and right images before they are fed into the depth estimation model, preventing pixel mismatch issues from occurring in the first place
Solution Approach 2:
The patent introduces an intermediary processing stage between image input and depth estimation. The autoencoder-based image restoration module and image fusion module act as intermediaries that process the raw binocular images, restoring degraded content and fusing corresponding pixels from both views to create consistent input for the depth estimation model
2Reliability
If image restoration and fusion processing is performed before depth estimation, then pixel value differences are reduced and model reliability improves, but processing complexity increases
Solution Approach 1:
The patent applies self-service by using an autoencoder-based unsupervised image restoration approach. The model learns to restore degraded images by training on pairs of degraded and restored images, enabling the system to automatically correct image quality issues without requiring manual intervention or complex supervised annotation processes
Solution Approach 2:
The patent merges multiple processing functions into integrated modules. The image restoration and image fusion operations are combined in a sequential processing pipeline that handles both degradation correction and pixel alignment together, reducing overall system complexity compared to separate independent processing stages
Data Source
AI summary
A method for training a depth estimation model is provided. The method includes obtaining a first left image and a first right image. A disparity map is obtained by inputting the first left image into a depth estimation model. A second right image is obtained by adding the first left image to the disparity map. The first left image is converted into a third right image. A mask image is obtained by performing a binarization processing on a pixel value of each of pixel points of the third right image. Once a loss value of the depth estimation model is obtained by calculating a mean square error of pixel values of all corresponding pixel points of the first right image, the second right image, and the mask image, a depth estimation model is iteratively trained according to the loss value.