Hybrid Dilated Convolution for Semantic Segmentation Gridding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semantic segmentation systems face issues such as 'gridding' due to standard dilated convolution operations and loss of information during downsampling, which affects the identification of important objects in images.
Innovation Solution
The implementation of dense upsampling convolution (DUC) for pixel-level prediction and a hybrid dilated convolution (HDC) framework that alleviates the 'gridding' issue and enlarges the receptive fields of the network, enabling more accurate semantic segmentation by aggregating global information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If standard dilated convolution operation is used to enlarge the field of convolutional kernels, then the receptive field is enlarged and global information is aggregated, but a 'gridding issue' is produced that degrades segmentation quality
Solution Approach 1:
The patent segments the convolution operation into two distinct phases: a sparse convolution phase that efficiently aggregates global information with large receptive field, and a dense convolution phase that refines local details and eliminates gridding artifacts. This segmentation allows each phase to specialize in different aspects of feature extraction, resolving the contradiction between global context aggregation and local precision.
Solution Approach 2:
The patent applies different convolution strategies to different spatial regions and scales: sparse convolution with large kernels for global context, and dense convolution with small kernels for local details. This local quality principle ensures that each region of the feature map receives appropriate processing tailored to its specific needs, eliminating gridding while preserving both global and local information.
2Productivity
If max-pooling or strided convolution is used to downsample feature maps, then computational efficiency is improved and memory usage is reduced, but information is lost that is necessary for identifying important objects
Solution Approach 1:
The patent performs preliminary dense convolution operations at earlier stages to extract and preserve fine-grained object details before downsampling occurs. This preliminary action ensures that critical object identification information is captured and retained in the feature hierarchy, preventing information loss during subsequent max-pooling or strided convolution operations.
Solution Approach 2:
The patent implements a nested architecture where dense convolutional layers are embedded within the downsampling pathway, creating multiple scales of feature representation. Each dense convolutional block operates at a different resolution level, with finer details preserved at deeper levels and coarser features at shallower levels, allowing information to be retained across multiple scales simultaneously.
3Measurement precision
If dense upsampling convolution is used to generate pixel-level prediction, then detailed information is recovered and pixel-level accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent applies dense upsampling convolution selectively rather than uniformly across all feature maps. It focuses computational resources on regions and scales where detailed information is most critical for segmentation accuracy, using partial dense convolution operations that recover fine details only where needed, thus balancing precision with computational complexity.
Data Source
AI summary
A system and method for semantic segmentation using hybrid dilated convolution (HDC) are disclosed. A particular embodiment includes: receiving an input image; producing a feature map from the input image; performing a convolution operation on the feature map and producing multiple convolution layers; grouping the multiple convolution layers into a plurality of groups; applying different dilation rates for different convolution layers in a single group of the plurality of groups; and applying a same dilation rate setting across all groups of the plurality of groups.