Semantic Segmentation Feature Maps With Multi-Scale Parallel Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semantic segmentation methods using Atrous Spatial Pyramid Pooling (ASPP) and pooling layers fail to utilize high-level feature maps effectively, leading to a loss of detail information and neglect of local information, resulting in rough object predictions.
Innovation Solution
A neural network architecture with parallel units containing cascaded convolution and pooling layers of varying kernel sizes, followed by an aggregation part to concatenate results, enhancing the utilization of high-level feature maps and maintaining object details.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Atrous Spatial Pyramid Pooling (ASPP) and pooling layers are used to process feature maps, then the model can capture multi-scale context information, but detail information is lost and local information is neglected
Solution Approach 1:
The patent divides the feature map processing into multiple parallel paths with different kernel sizes (1x1, 3x3, 5x5, 7x7 convolutions) and corresponding pooling operations. Each path processes feature maps at different scales independently, allowing the model to capture multi-scale context while preserving detailed information through the smallest kernel path that maintains spatial resolution.
Solution Approach 2:
The patent combines the outputs of multiple parallel paths (convolution paths and pooling paths) through concatenation and addition operations. This merging mechanism integrates multi-scale context information from different paths while preserving the complementary strengths of each path, including detailed local information from the 1x1 convolution path.
2Device complexity
If pooling layers are used to downsample feature maps, then the model can reduce computational complexity, but local information is neglected
Solution Approach 1:
The patent segments the feature map processing into multiple parallel paths where different paths perform different operations. The 1x1 convolution path maintains spatial resolution without downsampling, preserving local information, while other paths with larger kernels and pooling operations reduce complexity for capturing broader context. This segmentation allows the model to balance computational complexity with information preservation.
Solution Approach 2:
The patent applies different processing qualities to different regions of the feature map through parallel paths with different kernel sizes. The 1x1 convolution path provides high-resolution local detail, while paths with larger kernels (3x3, 5x5, 7x7) provide broader context. This local quality differentiation ensures that local information is preserved where needed while reducing complexity where broader context is more important.
Data Source
Figure 1
Figure 2a
Figure 2b
AI summary
Method and apparatus are disclosed for computer vision. The method may comprise processing, by using a neural network, input feature maps of an image to obtain output feature maps of the image. The neural network may comprise a convolution part and/or a pooling part, and an aggregation part. The convolution part may comprise at least one parallel unit each of which contains two parallel paths, each path of the two parallel paths contains two cascaded convolution layers. The kernel sizes are 1 dimension and are different in different unit. The pooling part comprises at least one parallel unit each of which contains two parallel paths, each path of the two parallel paths contains two cascaded pooling layers. The size of filters of pooling is 1 dimension and is different in different unit. The aggregation part is configured to concatenate results of the convolution part and/or the pooling part to obtain the output feature maps of the image.