CNN Image Segmentation Using Temporal Weight Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image segmentation technologies suffer from the dithering phenomenon in image edges due to performing segmentation on different image frames independently, which affects the accuracy and smoothness of object recognition in image understanding applications.
Innovation Solution
An image segmentation method utilizing a neural network where previous convolutional neural network (CNN) weight data is used to perform convolution on current images, generating smoother image segmentation results by incorporating time sequence information, thereby reducing dithering and enabling better object recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If image segmentation is performed independently on each image frame, then processing speed is maintained, but dithering phenomenon occurs at image edges
Solution Approach 1:
The patent applies preliminary action by using segmentation results from previous frames to generate initial segmentation results for the current frame. This allows the system to start with a head start based on historical data, improving both speed and accuracy by avoiding complete re-segmentation while maintaining edge smoothness through temporal consistency.
Solution Approach 2:
The patent implements continuity of useful action by maintaining temporal continuity in the segmentation process. Instead of treating each frame independently, the system continuously refines segmentation results across frames, ensuring that edge information is preserved and smoothed over time while maintaining processing efficiency.
2Device complexity
If image segmentation is performed on each frame independently, then computational complexity is reduced, but object recognition accuracy deteriorates due to dithering
Solution Approach 1:
The patent uses preliminary segmentation results from previous frames as initial estimates for the current frame. This preliminary action reduces the computational burden of complete re-segmentation while improving object recognition accuracy by providing a stable starting point that avoids dithering-induced errors.
Solution Approach 2:
The patent implements feedback by using segmentation results from previous frames to inform and refine the segmentation of the current frame. This feedback mechanism allows the system to correct dithering artifacts and improve object recognition accuracy without requiring significantly increased computational complexity.
3Ease of operation
If traditional segmentation methods are used, then processing simplicity is maintained, but edge smoothness and temporal consistency are poor
Solution Approach 1:
The patent maintains processing simplicity by building upon existing segmentation results rather than requiring complete re-processing. The preliminary action of using previous frame results as initial estimates simplifies the current frame processing while significantly improving temporal consistency and edge smoothness.
Solution Approach 2:
The patent achieves temporal consistency through continuity of useful action by carrying forward segmentation information across frames. This approach maintains the simplicity of individual frame processing while ensuring that object compositions remain stable and consistent over time, eliminating dithering effects.
Data Source
AI summary
An image segmentation method for performing image segmentation through a neural network implemented by an image segmentation apparatus is provided. The image segmentation method includes the steps outlined below. Previous CNN weight data is received by a current convolution neural network unit of the neural network, wherein the previous CNN weight data is generated by a previous convolution neural network unit of the neural network based on a previous image of video data corresponding to a previous time spot. A current image of the video data corresponding to a current time spot next to the previous time spot is received by the current convolution neural network unit. Convolution is performed according to the previous CNN weight data and the current image to generate a current image segmentation result by the current convolution neural network unit.