Semantic Segmentation via Multi-Branch Neural Network Averaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning methods for image semantic segmentation face challenges such as mismatching, fuzzy classification, and ignoring small objects, which affect the accuracy and reliability of pixel-level classification in images.
Innovation Solution
The method employs a neural network with a residual network and multiple feature analyzing branches, including dual attention networks, pyramid pooling modules, and conditional random fields, to extract and analyze feature maps, generating predictions for subpixels by averaging outputs from different branches, thereby improving segmentation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current deep learning methods are used for image semantic segmentation, then the basic pixel-level classification function is achieved, but mismatching and fuzzy classification occur reducing accuracy
Solution Approach 1:
The patent divides the feature analysis process into multiple independent branches (first feature analyzing branch with dual attention network, second feature analyzing branch with pyramid pooling module). Each branch processes features differently and generates separate predictions, which are then combined to produce the final segmentation result. This multi-branch segmentation approach resolves the contradiction by allowing each branch to specialize in different aspects of feature analysis, thereby improving both accuracy and reliability simultaneously.
Solution Approach 2:
The patent merges the predictions from multiple feature analyzing branches through averaging to generate the final segmentation prediction. By combining the outputs of the dual attention network branch and the pyramid pooling module branch, the system leverages the strengths of different feature analysis approaches, improving classification reliability while maintaining high accuracy across diverse segmentation scenarios.
2Productivity
If current deep learning methods are used for image semantic segmentation, then processing speed is maintained, but small objects are ignored reducing segmentation quality
Solution Approach 1:
The patent introduces pyramid pooling modules that analyze features at multiple scales and dimensions. By processing features through different pooling levels, the system captures both fine-grained details of small objects and broader contextual information, enabling accurate segmentation of small objects without significantly increasing processing time. This multi-dimensional feature analysis resolves the contradiction between processing speed and small object segmentation accuracy.
Solution Approach 2:
The dual attention network implements local quality enhancement by applying attention mechanisms that adaptively weight different regions of the feature maps. This allows the network to focus computational resources on regions containing small objects while maintaining efficient processing of other areas, thereby improving small object segmentation accuracy without proportionally increasing overall processing time.
3Measurement precision
If multiple feature analyzing branches are introduced to improve segmentation accuracy, then classification precision improves, but device complexity increases
Solution Approach 1:
The patent segments the complex feature analysis task into distinct, modular branches (dual attention network branch and pyramid pooling module branch). Each branch is independently designed and optimized for specific feature analysis functions, making the overall complex system more manageable and interpretable. This modular segmentation allows high segmentation accuracy to be achieved through coordinated simple components rather than a single complex network.
Solution Approach 2:
The patent designs feature analyzing branches that perform multiple functions: the dual attention network handles both spatial and channel attention, while the pyramid pooling module simultaneously performs feature extraction and multi-scale analysis. This multi-functionality reduces the need for additional specialized components, thereby improving segmentation accuracy without proportionally increasing device complexity.
Data Source
AI summary
A method of semantically segmenting an input image using a neural network is provided. The method includes extracting features of the input image to generate one or more feature maps; and analyzing the one or more feature maps to generate a plurality of predictions respectively corresponding to a plurality of subpixels of the input image. Extracting features of the input image is performed using a residual network having N number of residual blocks, N being a positive integer greater than 1. Analyzing the one or more feature maps is performed through M number of feature analyzing branches to generate M sets of predictions. A respective one set of the M sets of predictions includes multiple predictions respectively corresponding to the plurality of subpixels of the input image. A respective one of the plurality of predictions is an average value of corresponding ones of the M sets of predictions.