Hierarchical Neural Network Feature Integration for Efficient Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image recognition techniques using convolutional neural networks (CNNs) require high calculation costs when processing multiple candidate regions or performing multiple recognition tasks simultaneously, leading to inefficiencies in processing speed and accuracy.
Innovation Solution
The proposed solution involves an image processing apparatus and method that generates a connected layer feature by connecting outputs of multiple layers of a hierarchical neural network, which is then used to create attribute score maps for each region of an input image, allowing for efficient integration and output of recognition results without the need for high-cost processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If features are totaled for each region of interest using conventional CNN methods, then object detection can be performed, but calculation cost becomes high when there are many candidate regions or multiple recognition tasks
Solution Approach 1:
The patent segments the recognition process into two distinct phases: (1) a shared feature extraction phase that processes the entire image once to generate feature maps, and (2) multiple task-specific processing phases that operate independently on these feature maps. This segmentation allows different recognition tasks to share computational resources, significantly reducing calculation cost when multiple tasks are performed simultaneously while maintaining detection accuracy.
Solution Approach 2:
The patent performs preliminary feature extraction across the entire image before specific region-based processing. By pre-computing feature maps at multiple scales and orientations at the image level, the system avoids redundant calculations when processing multiple candidate regions or tasks, as all subsequent operations can directly utilize these pre-computed features.
2Adaptability or versatility
If conventional processing is used to handle multiple candidate regions or recognition tasks, then comprehensive recognition can be achieved, but processing speed decreases due to high calculation cost
Solution Approach 1:
The patent creates a universal feature representation that serves multiple recognition tasks simultaneously. The shared feature extraction module generates multi-scale and multi-orientation feature maps that can be applied to various recognition tasks (object detection, scene recognition, etc.) without requiring separate processing pipelines, thereby improving processing speed while maintaining multi-task capability.
Solution Approach 2:
The patent introduces additional dimensions to the feature representation by generating feature maps at multiple scales and orientations. This dimensional expansion allows the system to handle diverse recognition tasks more efficiently, as the enriched feature space provides sufficient information for various tasks without requiring redundant processing, thus improving processing speed.
3Measurement precision
If multiple layers of hierarchical neural network are processed separately for each region, then detailed feature extraction can be achieved, but calculation cost increases significantly
Solution Approach 1:
The patent merges the feature extraction operations across multiple layers and regions by implementing a shared hierarchical neural network that processes the entire image once. The feature maps generated at each layer are reused across different regions and tasks, reducing processing complexity while maintaining the detailed feature extraction capability provided by multiple network layers.
Data Source
AI summary
A connected layer feature is generated by connecting outputs of a plurality of layers of a hierarchical neural network obtained by processing an input image using the hierarchical neural network. An attribute score map representing an attribute of each region of the input image is generated for each attribute using the connected layer feature. A recognition result for a recognition target is generated and output by integrating the generated attribute score maps for respective attributes.