Embedded Target Detection with Shared Base Network and Self-Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current target detection systems for embedded devices face challenges such as high computational requirements, large storage needs, and limited capability for offline operation, especially in resource-constrained environments and varying lighting conditions.
Innovation Solution
A target detection system for embedded devices featuring a multi-layer structure with branch modules, a shared base network, a private base network, and a detection module, along with an online model self-calibration system that collects samples, annotates them, and updates model parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are used for target detection, then detection performance is improved, but computational requirements and storage space increase significantly
Solution Approach 1:
The network is segmented into a shared base network and multiple private base networks at different layers. The shared base network processes common features, while each private base network handles specific layer requirements. This segmentation allows the system to achieve deep learning performance while reducing overall computational burden by avoiding redundant processing in the shared portion.
Solution Approach 2:
The network architecture implements a nested structure where the shared base network is embedded within all private base networks. The output of the shared base network serves as input to multiple private base networks simultaneously. This nesting enables efficient feature reuse and reduces computational complexity while maintaining detection performance.
2Measurement precision
If deep learning models with many parameters are deployed, then detection capability is improved, but storage space requirements increase
Solution Approach 1:
Multiple base networks are merged into a single shared base network that is commonly used across all layers. This merging consolidates redundant parameters and enables parameter sharing, significantly reducing the total number of parameters required while maintaining the detection capabilities provided by deep learning.
3Productivity
If a small network model is used to reduce parameters and calculation, then offline detection is enabled, but detection rate decreases significantly in darker environments
Solution Approach 1:
The network architecture dynamically adjusts the number of MobileNet blocks in the shared base network based on target difficulty. For easier targets, fewer blocks are used, enabling fast offline detection. For difficult targets (such as those in darker environments), the network can access deeper layers with more blocks to improve detection rate, thus adapting performance to task requirements.
4Loss of time
If a small network model is trained with limited data, then training time is reduced, but the network fails to generalize to different background states
Solution Approach 1:
The shared base network is designed to serve multiple functions across different layers and detection tasks. It extracts universal features that can be reused by private base networks for various detection scenarios. This multi-functionality enables the network to generalize better to different background states and lighting conditions without requiring extensive retraining.
5Adaptability or versatility
If manual annotation of actual collected images is performed to train the network, then adaptation to real conditions is improved, but manpower and material resources increase significantly
Solution Approach 1:
The system implements self-service through automatic annotation functionality. The annotation module automatically processes collected images, identifies targets, and generates training data without requiring manual intervention. This self-service capability maintains adaptation to real conditions while dramatically reducing the manpower and resources required for training.
Data Source
AI summary
A target detection system suitable for an embedded device, comprising an embedded device (5) and a server (6); target detection logic (5.1) running in the embedded device (5) is composed of a multi-layer shared base network, a private base network, and a detection module; a parameter of the shared base network directly comes from an output of an upper layer; and an image is processed by the shared base network and the private base network to obtain a feature map, and after being processed by the detection module, a result merging module merges and outputs a target detection result. The target detection system further comprises an online model self-calibration system. After collecting a sample, the embedded device (5) irregularly uploads the sample to the server (6), and after labeling the sample by means of automatic and manual methods, the server (6) trains a model and updates same to the embedded device (5). The target detection system can perform well in an embedded device (5), uses a large-scale target detection model on a server (6) to complete automatic labeling which reduces workload, and completes model correction more efficiently.