Edge Device Object Detection Model Refinement via Cloud Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Edge devices face inefficiencies in object detection due to incomplete image data, leading to undesired classification and low accuracy rates, as models trained on small datasets tend to over-fit and incorrectly identify objects, such as distinguishing between fruits like oranges and lemons.
Innovation Solution
An edge device system that captures images, transmits them to a cloud server for object detection using CNN algorithms, displays found objects for user confirmation, and refines the model based on user feedback, incorporating additional images to improve accuracy, thereby reducing over-fitting and enhancing detection precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a model is trained on a small number of images captured by the edge device, then the training process is fast and resource-efficient, but the model accuracy is low and the model over-fits
Solution Approach 1:
The patent combines local image data captured by the edge device with additional images obtained from the cloud server to create a more comprehensive training dataset. This merging of data sources allows the model to be trained on more diverse and abundant images, improving accuracy and reducing over-fitting while maintaining the efficiency of local processing.
Solution Approach 2:
The cloud server acts as an intermediary that supplements the limited local image data with additional relevant images. This intermediary resource enables the system to overcome the constraint of small local datasets without requiring the edge device to capture excessive images, thus balancing training quality and efficiency.
2Device complexity
If the model is trained with incomplete image data, then the training process is simple and quick, but the model produces undesired classification and detection results
Solution Approach 1:
The system implements a feedback mechanism where the model's detection results are evaluated, and additional images are acquired based on the identified deficiencies. The model is then retrained with this supplementary data, creating an iterative improvement loop that enhances classification reliability while managing training complexity through targeted data acquisition.
Solution Approach 2:
The system performs preliminary object detection on the initial image set to identify what types of objects are present and what additional data would be beneficial. This preliminary analysis guides the subsequent acquisition of supplementary images, ensuring that the additional data directly addresses the specific deficiencies in the initial dataset.
3Measurement precision
If multiple training iterations are performed to improve model accuracy, then the detection precision improves, but the time and resources required increase significantly
Solution Approach 1:
Instead of performing multiple complete training iterations from scratch, the system acquires additional images selectively based on the specific deficiencies identified in the initial model performance. This partial action approach targets only the necessary improvements, achieving better detection precision without the time cost of repeated full training cycles.
Data Source
AI summary
A method for detecting a specific object based on a specific model includes: capturing a set of images, wherein objects in each image include a desired object; transmitting the set of images to a cloud server; in response to found objects being obtained from the set of images based on at least one object detection algorithm in the cloud server, displaying the found objects fora user to confirm which object is desired; in response to the desired object being confirmed from the found objects that are displayed, loading the specific model of the desired object from the cloud server, wherein the specific model of the desired object is trained on the cloud server based on at least the set of images and related CNN algorithm; and performing the specific model to detect the specific object on a captured image.