Scene Parsing via Multi-Scale Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current scene parsing technologies face challenges in accurately categorizing pixels in images, especially when surrounding information is complex, leading to prediction failures due to similar colors and textures.
Innovation Solution
A method involving a deep neural network that performs convolution and pooling operations to generate feature maps, considering both semantic and spatial context information, forming a pyramid structure to enhance pixel categorization accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional scene parsing methods are used, then the processing speed is fast, but the prediction accuracy fails when surrounding information is complex
Solution Approach 1:
The patent segments the feature extraction process into multiple parallel convolutional branches, each processing different spatial receptive fields independently. This segmentation allows the system to capture both local and global context information simultaneously, improving pixel categorization accuracy while maintaining manageable computational complexity through modular architecture design.
Solution Approach 2:
The patent introduces a dimensional transformation by processing features at multiple spatial scales (different receptive fields) simultaneously. By stacking feature maps from different convolutional layers with varying kernel sizes, the system creates a multi-dimensional feature representation that captures context information across different spatial dimensions, thereby improving accuracy without linearly increasing complexity.
2Reliability
If multiple convolutional layers with large kernels are used, then context information is captured better, but computational complexity increases
Solution Approach 1:
The patent divides the computational workload into multiple parallel convolutional branches, each handling a specific receptive field size. This segmentation allows the system to distribute computational energy consumption across multiple smaller operations rather than one large computationally intensive operation, improving prediction reliability through comprehensive context capture while managing energy usage through parallel processing efficiency.
Solution Approach 2:
The patent employs multiple convolutional layers with progressively larger kernels to capture context information at different scales. By using partial actions (smaller kernels for local context, larger kernels for global context) rather than a single excessive operation, the system achieves reliable predictions through cumulative context information while optimizing computational energy consumption through efficient multi-scale processing.
Data Source
AI summary
A method for scene parsing includes: performing a convolution operation on a to-be-parsed image by using a deep neural network to obtain a first feature map, the first feature map including features of at least one pixel in the image; performing a pooling operation on the first feature map to obtain at least one second feature map, a size of the second feature map being less than that of the first feature map; and performing scene parsing on the image according to the first feature map and the at least one second feature map to obtain a scene parsing result of the image, the scene parsing result including a category of the at least one pixel in the image. A system for scene parsing and a non-transitory computer-readable storage medium can facilitate realizing the method.