Image Segmentation Using Resolution Transformation Matrix
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image semantic segmentation methods using convolutional neural networks (CNNs) face high calculation amounts and memory overheads due to the need to fuse feature maps with high resolutions, which limits their efficiency and scalability in applications like autonomous driving and augmented reality.
Innovation Solution
The proposed method involves obtaining an input image, performing multi-layer feature extraction to downsample feature maps to a reference resolution, fusing these maps, and then upsampling them using a transformation matrix W to achieve the original image resolution, thereby reducing computational and memory requirements while maintaining segmentation precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature maps with high resolutions are fused using conventional CNN methods, then segmentation precision is improved, but calculation amount and memory overheads increase significantly
Solution Approach 1:
The patent segments the feature processing into two distinct stages: (1) downsample feature maps to a reference resolution before fusion to reduce memory overhead, and (2) upsample the fused feature map back to original resolution for precise segmentation. This segmentation of the processing pipeline allows low-resolution fusion (saving memory) while still achieving high-resolution segmentation results.
Solution Approach 2:
The patent introduces a resolution dimension transformation by downsampling feature maps to a reference resolution for fusion operations, then upsampling the result back to the original resolution. This dimensional transformation allows the system to perform fusion operations on smaller data structures (reducing memory overhead) while maintaining the ability to produce high-resolution segmentation outputs.
2Measurement precision
If feature maps with high resolutions are fused using conventional CNN methods, then segmentation precision is improved, but calculation amount increases significantly
Solution Approach 1:
The patent segments the feature processing into two distinct stages: (1) downsample feature maps to a reference resolution before fusion to reduce memory overhead, and (2) upsample the fused feature map back to original resolution for precise segmentation. This segmentation of the processing pipeline allows low-resolution fusion (saving memory) while still achieving high-resolution segmentation results.
Solution Approach 2:
The patent introduces a resolution dimension transformation by downsampling feature maps to a reference resolution for fusion operations, then upsampling the result back to the original resolution. This dimensional transformation allows the system to perform fusion operations on smaller data structures (reducing memory overhead) while maintaining the ability to produce high-resolution segmentation outputs.
3Productivity
If feature maps are downsampled to reference resolution before fusion, then memory overhead and calculation amount are reduced, but segmentation precision may deteriorate
Solution Approach 1:
The patent performs preliminary downsampling of feature maps to reference resolution before fusion operations to reduce computational burden. However, it preserves the ability to restore full resolution through upsampling after fusion, ensuring that the final segmentation maintains high precision despite the intermediate low-resolution processing.
Solution Approach 2:
The patent dynamically changes the resolution parameter of feature maps during processing: downsampling to reference resolution for fusion operations (improving efficiency), then upsampling back to original resolution for final segmentation (maintaining precision). This parameter transformation allows the system to optimize both computational efficiency and segmentation accuracy at different stages of the pipeline.
Data Source
AI summary
This application discloses an image segmentation method in the field of artificial intelligence. The method includes: obtaining an input image and a processing requirement; performing multi-layer feature extraction on the input image to obtain a plurality of feature maps; downsampling the plurality of feature maps to obtain a plurality of feature maps with a reference resolution, where the reference resolution is less than a resolution of the input image; fusing the plurality of feature maps with the reference resolution to obtain at least one feature map group; upsampling the feature map group by using a transformation matrix W, to obtain a target feature map group; and performing target processing on the target feature map group based on the processing requirement to obtain a target image.