Distance Image Noise Reduction via Adaptive CNN Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image noise removal methods using convolutional neural networks (CNNs) lack control over smoothing strength, leading to residual noise in regions with high noise and oversmoothing in regions with low noise, which deteriorates noise reduction accuracy.
Innovation Solution
A processing circuit that inputs both the noise reduction target image and a noise correlation map to a CNN, allowing for noise reduction with intensity corresponding to the noise amount in each region, using a learned model to generate a denoise image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a CNN is used for noise removal with fixed convolution weights, then the process is simple and fast, but noise remains in high-noise regions and original signal is smoothed in low-noise regions
Solution Approach 1:
The patent applies dynamics by making the convolution weights variable rather than fixed. The weights are dynamically adjusted based on the noise amount detected in each region through the feedback loop, allowing the system to adapt to different noise conditions and achieve both speed and accuracy
Solution Approach 2:
The patent implements feedback by introducing a noise amount detection unit that continuously monitors the noise level in each region and feeds this information back to the convolution weight adjustment unit. This closed-loop feedback mechanism enables real-time optimization of noise removal performance
2Measurement precision
If smoothing strength is increased to remove noise in high-noise regions, then noise is reduced, but original signal is过度smoothed in low-noise regions
Solution Approach 1:
The patent applies local quality by adjusting the convolution weights individually for each region based on its specific noise characteristics. High-noise regions receive stronger smoothing weights while low-noise regions receive weaker weights, allowing customized noise removal that preserves signal integrity in each area
Solution Approach 2:
The patent implements parameter changes by dynamically modifying the convolution weights as a function of the detected noise amount. The weight parameter is changed from a fixed value to a variable value that adapts to local noise conditions, enabling optimal noise removal without excessive smoothing
Data Source
AI summary
According to one embodiment, a distance measuring apparatus includes a processing circuit and a memory. The processing circuit generates a distance image by measuring distances to an obstacle by using the time until the emitted light is reflected from the obstacle and returned as reflected light. The processing circuit generates an intensity image by measuring the intensity of the reflected light or the intensity of the environmental light. The memory stores a learned model for generating a denoise image, in which noise of the distance image is reduced, based on the distance image and the intensity image. The processing circuit inputs the distance image and the intensity image to the learned model, and generates the