Branching Identification Layers in Deep Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems face challenges in efficiently performing recognition tasks across various scales, as they struggle to divert feature amounts from intermediate layers and branch identification layers effectively for multiple recognition tasks within a single network structure.
Innovation Solution
The proposed solution involves a Deep Neural Network (DNN) with a multilayer Bidirectional Long Short-Term Memory (BLSTM) architecture, where identification layers are branched from a single intermediate layer to perform recognition tasks at different scales, such as action, play, and event segment recognition in video analysis, and edge detection, object contour, and object detection in image analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single neural network structure is used to perform multiple recognition tasks with different scales, then the device complexity is reduced, but it becomes difficult to effectively divert feature amounts from intermediate layers and branch identification layers for each task
Solution Approach 1:
The patent segments the neural network into multiple independent identification layers (first identification layer, second identification layer, etc.) that branch from a shared intermediate layer. Each identification layer is independently configured and optimized for specific recognition tasks at different scales, allowing the system to handle multiple tasks simultaneously while maintaining structural efficiency through the shared intermediate feature extraction.
2Productivity
If identification layers are branched from one intermediate layer for multiple recognition tasks, then the system can perform concurrent recognition at different scales, but it becomes difficult to optimize parameters for each task independently
Solution Approach 1:
The patent applies local quality by allowing each identification layer to have independently optimized parameters and configurations tailored to its specific recognition task. The first identification layer can be optimized for small-scale features while the second identification layer is optimized for large-scale features, with each layer having its own learnable parameters that are updated independently during training based on task-specific loss functions.
3Measurement precision
If separate neural networks are used for recognition tasks with different scales, then each task can be optimized independently, but the overall device complexity and computational overhead increase
Solution Approach 1:
The patent merges multiple recognition functions into a single unified neural network structure where different identification layers share a common intermediate feature extraction layer. This allows the system to maintain the recognition accuracy of specialized networks while reducing overall complexity through parameter sharing and joint training, as the intermediate layers extract features that are useful for multiple tasks simultaneously.
Data Source
AI summary
A Deep Neural Network (DNN) having a plurality of recognition tasks with different scales makes it possible to perform recognition processing in a network where identification layers are branched from one intermediate layer. An image recognition apparatus for recognizing a target includes a first acquisition unit configured to acquire from an input image a first intermediate feature amount for performing first identification, a first identification unit configured to perform the first identification based on the first intermediate feature amount, a second acquisition unit configured to acquire from the first intermediate feature amount a second intermediate feature amount for performing second identification having a larger scale than the first identification, and a second identification unit configured to perform the second identification based on the second intermediate feature amount.