Multi-Task Target Recognition Network for Low-Memory Video Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video monitoring systems face challenges with high memory occupation, bandwidth load, and long training duration due to independent-task algorithms for target detection and recognition, which are not suitable for scenes with limited resources and real-time requirements.
Innovation Solution
A multi-task network model is used for target detection and recognition, incorporating a feature extraction module, multi-scale feature fusion, and a detection head module with shared convolution layers, reducing the need for cloud server calls and optimizing memory and bandwidth usage by performing local image quality evaluation and cropping target images before recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If independent-task algorithms are used for target detection and recognition, then recognition accuracy is improved, but model training duration and camera side memory occupation increase
Solution Approach 1:
The patent merges target detection and target recognition into a single multi-task network model, where both functions are performed simultaneously within one model framework. This integration reduces the need for separate independent models, thereby decreasing overall training duration while maintaining recognition accuracy through shared feature extraction capabilities.
Solution Approach 2:
The multi-task network model is designed to perform multiple functions (target detection and target recognition) within a single universal framework. The model uses shared convolutional layers for feature extraction that serve both detection and recognition tasks, reducing redundant computation and training time while maintaining high accuracy for both functions.
2Measurement precision
If independent image quality evaluation model is used, then misrecognition rate is reduced, but camera side memory occupation and bandwidth load increase
Solution Approach 1:
The patent integrates image quality evaluation as an additional task within the multi-task network model, combining it with target detection and recognition. This integration allows the system to share computational resources and memory across multiple functions, reducing the overall memory footprint on the camera side while maintaining the ability to evaluate image quality for reducing misrecognition rates.
3Quantity of substance
If target image quality evaluation is integrated into target recognition algorithm, then camera side memory occupation is reduced, but transmission cost of target images to cloud server increases
Solution Approach 1:
The patent performs image quality evaluation as a preliminary action within the multi-task network model before target recognition. By evaluating image quality early in the processing pipeline, the system can identify and filter out low-quality images that would not contribute meaningfully to recognition results, thereby reducing the number of images that need to be transmitted to the cloud server and lowering transmission costs.
Solution Approach 2:
The multi-task network model performs image quality evaluation locally on the camera side using shared computational resources, enabling the system to self-assess image quality without requiring additional external evaluation systems. This local self-service capability reduces memory occupation while intelligently filtering images to minimize transmission costs to the cloud server.
Data Source
AI summary
This application provides target recognition method, a multi-task network model training method, and an electronic device. The target recognition method includes: inputting video images into a multi-task network model one by one to obtain a predicted feature map; performing post-processing on the predicted feature map to obtain a target detection result; judging whether a target class confidence degree is greater than a preset confidence degree; if so, judging whether a target image quality score is greater than a preset score; if so, cropping out a target image from the video images according to a target detection box; and inputting the target image into a target recognition model corresponding to the target class to obtain a target name. In this way, this application decreases the number of calls of the recognition model, and also reduces a training duration of the model.