Video Noise Reduction Using Spatial-Temporal Patch Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise filtering technologies fail to effectively eliminate structural fixed pattern noise in video streams, particularly in thermal cameras, where noise is prominent due to low signal-to-noise levels and is exacerbated by camera movement, leading to distracting visual artifacts.
Innovation Solution
A method that compares pixel patches within a search area in one image frame to a target pixel patch, using a first threshold, and incorporates directional camera data to align with a previous image frame, allowing for weighted averaging to reduce noise while preserving edges, utilizing both spatial and temporal data for precise noise reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing noise filtering technologies are used, then processing speed is maintained, but structural fixed pattern noise is not effectively eliminated
Solution Approach 1:
The patent divides the image processing task into multiple passes: a first pass processes pixels using spatial information from the current frame, while a second pass uses temporal information from previous frames. This segmentation allows complex noise reduction to be broken into manageable steps that can be optimized for speed.
Solution Approach 2:
The patent performs preliminary noise reduction in the first pass using only spatial information before proceeding to the second pass with temporal information. This preliminary action reduces the overall noise level early, making subsequent processing more efficient and effective.
2Reliability
If noise filtering is applied to reduce fixed pattern noise, then image quality improves, but edge preservation becomes difficult
Solution Approach 1:
The patent applies different processing strategies to different regions of the image based on local characteristics. By analyzing pixel patches and their similarities, the method adapts the noise reduction strength locally, preserving edges in regions where they exist while aggressively reducing noise in homogeneous areas.
Solution Approach 2:
The patent uses dynamic thresholding and adaptive weighting in the averaging process. The similarity thresholds and weights applied to different pixel patches are adjusted dynamically based on local image characteristics, allowing the algorithm to preserve sharp edges while reducing noise in smoother regions.
3Reliability
If camera movement is compensated for, then fixed pattern noise reduction is improved, but processing complexity increases
Solution Approach 1:
The patent introduces camera directional data (from gyroscopes or accelerometers) as an intermediary to bridge the gap between actual camera movement and image frame alignment. This external sensor data serves as a mediator that simplifies the complexity of tracking and compensating for camera movements without requiring complex image analysis.
Solution Approach 2:
The patent replaces complex mechanical image stabilization systems with computational methods that use sensor data to guide the noise reduction process. Instead of physically stabilizing the camera mechanism, the system uses digital processing guided by inertial sensor measurements to achieve similar effects.
4Reliability
If spatial and temporal data are combined for noise reduction, then noise elimination is improved, but computational load increases
Solution Approach 1:
The patent segments the computational task into two distinct passes: spatial processing in the first pass and temporal processing in the second pass. This segmentation allows the system to manage computational energy by breaking down the complex task of combining spatial and temporal data into smaller, more energy-efficient steps.
Solution Approach 2:
The patent performs preliminary spatial noise reduction before applying temporal filtering. This preliminary action reduces the overall noise level and simplifies the subsequent temporal processing, reducing the computational energy required for the combined spatial-temporal analysis.
Data Source
Figure 1~5
Figure 2A~4A
Figure 2B~4B
AI summary
A method for structural fixed pattern noise reduction in a video stream comprises: for each target pixel (116) in a first image frame (114) defining a first target pixel patch (118) including the target pixel (116), defining a first search area (120) comprising the first target pixel patch (118), for each pixel in the first search area comparing a first pixel (124) patch around the pixel with the first target pixel patch (118), using the pixel when calulating an average for the target pixel (118) if the similarity of the pixel patch (124) and the first target pixel patch (118) is within a first threshold, and, in a second image frame (114') localizing a second target pixel (114'), defining a second pixel seach area (120') comprising the second target pixel (116'), for each pixel in the second search area (120') comparing a second pixel patch (124') around the pixel of the second search area (120') with the first target pixel patch (116), using the pixel of the second search area (124') when calculating an average of the first target pixel (116) if the similarity of the second pixel patch (124') and the first target pixel patch (118) is within a second threshold, wherein the first image frame (114) and the second image frame (114') are separate image frames in the videostream, and wherein the location of the second target pixel is estimated from the location of the first target pixel(116), combined with camera directional data.