Joint Perception Model Training with Shared Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision and deep learning technologies face challenges in efficiently performing multi-task joint perception tasks, such as target detection, attribute recognition, and search feature extraction, due to high computational requirements and resource inefficiencies.
Innovation Solution
A joint perception model training method that includes a feature extraction network and a joint perception network, where sample images are processed to obtain target sample features, and these features are used to perform joint perception and train the model, reducing the need for separate feature extraction networks for each perception task.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate feature extraction networks are used for each perception task, then each task can be optimized independently, but the computational requirements and resource consumption increase significantly
Solution Approach 1:
The patent merges multiple task-specific feature extraction networks into a single shared feature extraction network. This shared network extracts features that are common across multiple perception tasks (target detection, attribute recognition, search feature extraction), reducing redundant computations. The merged architecture processes images once and distributes the extracted features to multiple task-specific processing paths, significantly lowering computational resource consumption while maintaining task accuracy.
Solution Approach 2:
The shared feature extraction network is designed to serve multiple perception tasks simultaneously. Instead of having dedicated networks for each task, this universal network extracts general-purpose features that can be utilized by different downstream tasks. The network structure includes multiple output paths that can accommodate different perception tasks, making it multi-functional and adaptable to various analysis requirements.
2Reliability
If multiple separate perception networks are deployed for different tasks, then each task receives dedicated processing, but the device complexity and data operation amount increase
Solution Approach 1:
The patent segments the perception system into a shared feature extraction component and multiple task-specific processing components. The feature extraction network is segmented from the task execution networks, allowing features to be extracted once and then distributed to multiple tasks. This segmentation reduces overall system complexity by eliminating redundant feature extraction operations while maintaining dedicated processing paths for each task through the shared features.
Solution Approach 2:
The shared feature extraction network acts as an intermediary between the input images and the multiple task-specific processing networks. Instead of each task network directly processing raw images, they all receive pre-processed features from the intermediary feature extraction network. This intermediary structure simplifies the overall system by centralizing the complex feature extraction operations while allowing simpler task-specific networks to focus on their respective tasks.
3Measurement precision
If independent feature extraction is performed for each perception task, then task-specific features are optimized, but the data operation amount and computation time increase
Solution Approach 1:
The patent performs feature extraction as a preliminary action before task-specific processing. The shared feature extraction network extracts general features from images in advance, and these pre-extracted features are then fed to multiple task-specific processing paths. This preliminary feature extraction eliminates the need for each task to independently extract features from raw images, significantly improving computation efficiency while maintaining the ability to extract task-relevant features through the shared representation.
Data Source
AI summary
Provided are a joint perception model training method, a joint perception method, a device, and a storage medium. The joint perception model training method includes: acquiring sample images and perception tags of the sample images; acquiring a preset joint perception model, where the joint perception model includes a feature extraction network and a joint perception network; performing feature extraction on the sample images through the feature extraction network to obtain target sample features; performing joint perception through the joint perception network according to the target sample features to obtain perception prediction results; and training the preset joint perception model according to the perception prediction results and the perception tags, where the joint perception includes executing at least two perception tasks.