Mobile Object Detection via Segmented Neural Network Backbone and Head
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection models face challenges in achieving real-time performance and high accuracy on mobile devices, as they are often migrated from server-side models with compressed parameters, leading to a significant drop in accuracy and difficulty in meeting deployment requirements on mobile terminals.
Innovation Solution
A neural network model with sequentially connected feature extracting networks that perform multi-scale feature fusion, using skip connections and lightweight architectures like residual convolutional networks and depthwise separable convolutional layers, to enhance feature extraction and detection accuracy on mobile devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If server-side object detection models are migrated to mobile devices with compressed parameters, then deployment feasibility is improved, but detection accuracy deteriorates significantly
Solution Approach 1:
The model is segmented into two distinct parts: a large-scale backbone network trained on servers for comprehensive feature extraction, and a lightweight head network deployed on mobile devices for real-time detection. This segmentation allows the mobile device to only process essential detection tasks while relying on pre-trained backbone features, thus maintaining accuracy while ensuring deployment feasibility.
Solution Approach 2:
The backbone network is pre-trained on server-side with large-scale datasets and complex architectures before deployment. The pre-trained backbone extracts robust features that are then transferred to the mobile device, eliminating the need for the mobile device to perform heavy pre-training computations while preserving the benefits of extensive pre-training for accurate detection.
2Measurement precision
If complex two-stage object detection algorithms are deployed, then detection accuracy is improved, but prediction speed deteriorates
Solution Approach 1:
The detection pipeline is segmented into offline backbone training (accuracy-focused) and online head inference (speed-focused). The backbone network performs comprehensive feature extraction during offline training, while the lightweight head network performs rapid detection during online inference, combining the accuracy benefits of complex models with the speed of simple models.
Solution Approach 2:
The mobile device performs only the essential detection task through the lightweight head network, rather than executing the complete complex two-stage algorithm. This partial action approach achieves sufficient detection accuracy for mobile applications while dramatically reducing computational overhead and improving prediction speed.
3Productivity
If lightweight architectures are used to reduce parameter quantity, then computational efficiency is improved, but feature extraction capability deteriorates
Solution Approach 1:
The architecture is segmented into a large-scale backbone for comprehensive feature extraction and a lightweight head for efficient processing. The backbone, despite its size, is deployed once during offline training to extract robust features, while the lightweight head ensures high computational efficiency during mobile deployment, thus combining strong feature extraction with high efficiency.
Solution Approach 2:
The pre-trained backbone features act as an intermediary between the large-scale server model and the lightweight mobile model. These intermediate features capture comprehensive visual information and are transferred to the lightweight head, enabling the mobile device to achieve strong feature extraction capability without requiring a heavy architecture.
Data Source
AI summary
A method for detecting an object in an image includes: obtaining an image to be detected; generating a plurality of feature maps based on the image to be detected by a plurality of feature extracting networks in a neural network model trained for object detection, in which the plurality of feature extracting networks are connected sequentially, and input data of a latter feature extracting network in the plurality of feature extracting networks is based on output data and input data of a previous feature extracting network; and generating an object detection result based on the plurality of feature maps by an object detecting network in the neural network model.