Surface Identification via Embedding and Reconstruction Error
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to reliably identify the surface in images while accurately distinguishing it from anomalies present on the surface, particularly in the context of autonomous driving.
Innovation Solution
A computer-implemented method using a neural network-based embedding module to separate surface and non-surface pixels by configuring embedding vectors, employing a spatial pooling function and reconstruction techniques to enhance separation, and utilizing a loss function to train the system for accurate identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image analysis methods are used to identify surface pixels, then the processing is simpler and faster, but the reliability of distinguishing surface pixels from anomalies is insufficient
Solution Approach 1:
The patent applies segmentation by dividing the image processing task into multiple specialized modules: an embedding module that transforms pixel data into embedding vectors, a reconstruction module that generates reconstructed images, and an anomaly detection module that identifies discrepancies. This segmentation allows each module to specialize in a specific aspect of the problem, improving overall reliability while managing complexity through modular design.
Solution Approach 2:
The patent introduces embedding vectors as an intermediary representation between the original image pixels and the final anomaly detection. These embedding vectors serve as a mediator that captures essential surface characteristics in a transformed space, enabling more reliable distinction between surface pixels and anomalies while providing a structured intermediate step for processing.
2Measurement precision
If simple pixel classification is used, then the processing speed is faster, but the measurement precision in distinguishing surface from non-surface pixels deteriorates
Solution Approach 1:
The patent transforms the pixel classification problem from the original image space into an embedding space through the embedding module. This dimensionality change allows the system to analyze pixels in a transformed feature space where surface and non-surface characteristics are more distinctly separated, improving measurement precision without requiring exhaustive analysis of the original high-dimensional pixel data.
Solution Approach 2:
The embedding module performs preliminary transformation of pixel data into embedding vectors before the actual classification and anomaly detection takes place. This preliminary action pre-processes the data into a form that is more amenable to accurate classification, enabling faster and more precise processing in subsequent stages by working with compressed, meaningful representations rather than raw pixel values.
3Loss of information
If anomaly detection is performed using existing methods, then the system is simpler, but the loss of information in distinguishing anomalies from surface pixels increases
Solution Approach 1:
The patent implements a feedback mechanism through the reconstruction process. The reconstruction module generates reconstructed images based on the embedding vectors, and the anomaly detection module compares these reconstructed images with the original images. This feedback loop allows the system to continuously refine its understanding of surface characteristics versus anomalies, retaining more information about both by verifying predictions against reconstructed data.
Solution Approach 2:
The patent creates a copy of the original image through the reconstruction process. The reconstruction module generates a reconstructed version of the image from the embedding vectors, allowing the system to compare the original and reconstructed copies to identify anomalies. This copying approach preserves information about both the original surface characteristics and the detected anomalies by maintaining a reference copy for comparison.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A surface identification method for identifying the pixels of a scene image (si) which represent a specific surface, comprising: S22) encoding the scene image (si) into a segmentation latent features map (slfm); S24) based on that map (slfm), predicting a segmentation (sl,sm) classifying the pixels in surface or non-surface categorie(s); S30) based on the segmentation latent features map (slfm), calculating a segmentation embedding (se) such that the non-surface embedding vectors (nsev) are separated from the cluster (c) of the surface embedding vectors (sve); S42) based on the segmentation embedding (se), calculating a reconstructed scene image (rsi); S44) calculating a reconstruction error map (rerr) of the errors (err) between the reconstructed scene image (rsi) and the scene image (si); S70) based on the segmentation (sl,sm) and the reconstruction error map (rerr), calculating a surface pixels map (spm) in which the surface pixels of the scene image (ri) can be identified.