Hybrid Image Boundary Detection for Noisy Low-Light Scenes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image boundary detection methods struggle with accuracy in noisy images, especially those with low light levels, and are computationally inefficient, particularly in deep learning-based systems like Vision Transformer.
Innovation Solution
A two-stage hybrid neural network architecture combining convolutional neural networks (CNN) for initial boundary estimation and a feedforward transformer encoder for refinement, using field-of-junctions (FoJ) representations to enhance boundary detection accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning-based boundary detection methods are used, then boundary detection accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the image processing task into two distinct stages: a CNN-based initialization stage for local boundary detection and a Transformer-based refinement stage for global boundary optimization. This segmentation allows each stage to specialize in specific functions, reducing overall computational complexity while maintaining high accuracy.
Solution Approach 2:
The patent applies preliminary action by using the CNN initialization stage to perform rapid local boundary detection before the Transformer refinement stage processes the results. This preliminary processing provides a good initial estimate that reduces the computational burden on the refinement stage, thereby reducing overall computational complexity while maintaining accuracy.
2Measurement precision
If Vision Transformer model is used for boundary detection, then boundary detection accuracy is improved, but processing speed decreases
Solution Approach 1:
The patent segments the processing task by using a lightweight CNN for initial processing and reserving the more computationally intensive Transformer for refinement only. This segmentation ensures that the Transformer processes only processed data rather than raw images, significantly improving processing speed while maintaining accuracy.
Solution Approach 2:
The CNN initialization stage performs preliminary boundary detection quickly and accurately for local regions, providing a foundation that reduces the amount of data the Transformer needs to process. This preliminary action significantly speeds up the overall processing time while maintaining high accuracy through the subsequent refinement stage.
3Speed
If local boundary detection methods are used, then processing speed is maintained, but boundary detection accuracy deteriorates in noisy images
Solution Approach 1:
The patent merges the advantages of both CNN-based local detection and Transformer-based global refinement into a unified two-stage system. The CNN stage provides fast local detection while the Transformer stage adds global context and refinement, combining the speed and accuracy benefits of both approaches.
Solution Approach 2:
The patent uses the CNN initialization stage as an intermediary that processes images quickly and provides a good initial estimate. This intermediary then feeds processed data to the Transformer refinement stage, allowing the system to maintain speed while achieving high accuracy through the intermediary's preliminary processing.
Data Source
AI summary
Digital image processing with automated image boundary detection. A two-stage hybrid computer-implemented neural network architecture for a digital image processor includes an initialization stage having a convolutional neural network architecture and a refinement stage having a feedforward transformer encoder. The convolutional neural network architecture generates initial field-of-junctions parameters for overlapping image patches of an input image. The feedforward transformer encoder refines each of the initial field-of-junctions parameters to output refined field-of-junctions parameters for each of the image patches and generate a boundary map and a color map for each image patch. A multi-stage training scheme is used to optimize the parameters of the neural network architecture. The initialization stage is trained using the patch reconstruction loss. Then, the refinement stage is optimized by using a mean squared error loss function to directly supervise the initial field-of-junctions parameters, and then using a comprehensive image reconstruction loss to evaluate the loss in a single step.