Unified CNN with Shared Core for Multi-Category Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks (CNNs) are limited in simultaneously detecting multiple categories of elements of interest, such as people and vehicles, due to the requirement for fully annotated learning bases and the inability to merge partially annotated databases, leading to increased computational resources and manual annotation burdens.
Innovation Solution
A method for training a single CNN that can detect multiple categories by using a common core architecture with specific encoding layers for each training image base, allowing simultaneous learning from multiple databases without overlapping annotations, and employing distinct cost functions for each base to avoid category conflicts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple separate CNNs are trained on different annotated databases to detect multiple categories, then detection coverage for different categories is improved, but device complexity and computational resources increase
Solution Approach 1:
The patent merges multiple separate CNN detectors into a single unified CNN that can detect multiple categories simultaneously. The unified CNN shares common feature extraction layers and only diverges in the final classification layers, allowing it to process multiple category types (people, faces, vehicles) with a single model instance, thereby reducing device complexity while maintaining detection coverage
Solution Approach 2:
The unified CNN is designed with multi-functionality to handle multiple detection categories within a single model. By incorporating multiple encoding layers that can be selectively activated based on the input image characteristics, the single CNN achieves universal detection capability across different categories without requiring separate specialized detectors for each category
2Measurement precision
If multiple separate CNNs are used to detect multiple categories, then category-specific detection accuracy is improved, but processing time increases
Solution Approach 1:
The unified CNN employs segmentation of its architecture into a shared common core and multiple category-specific encoding layers. This allows the network to process features once in the common core and then branch into specific category detections, avoiding redundant processing that would occur with multiple separate CNNs, thereby maintaining accuracy while reducing processing time
3Adaptability or versatility
If training datasets are merged to include multiple categories, then comprehensive detection capability is improved, but annotation complexity and cost increase
Solution Approach 1:
The training approach segments the training process into separate phases for different categories. Each category is trained on its own annotated database independently, and the unified CNN learns to switch between different encoding layers based on the input characteristics. This segmentation allows using existing specialized annotated databases without requiring comprehensive multi-category annotations for all images, reducing annotation complexity
Solution Approach 2:
The patent introduces an intermediary mechanism in the form of multiple encoding layers that act as mediators between the common feature extraction core and the final classification outputs. These encoding layers enable the network to adapt to different category types without requiring the training data to be pre-annotated with all possible categories, allowing flexible use of existing annotated databases
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a method for learning parameters of a convolutional neural network (CNN) using data processing means (11a, 11b, 11c) from at least one server (1a, 1b, 1c), for detecting visible features of interest in images. The method is characterized in that it is implemented using a plurality of training image sets in which said features of interest are already annotated. The CNN is common to said plurality of training image sets and has a common core and a plurality of encoding layers, each specific to one of said plurality of training image sets. The present invention also relates to a method for detecting visible features of interest in an image.