Class-Focused Segmentation Model Training for Edge Deployment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation models are not suitable for on-device applications due to their large parameter count and resource requirements, and they struggle with the diversity of agricultural domains, leading to delayed deployment and poor performance when trained with limited data.
Innovation Solution
A two-stage model training approach is employed, first creating a master ML segmentation model and then a class-focused segmentation model using a reduced search space and hyperparameters, followed by fine-tuning on task-specific datasets to generate a compact and accurate model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing image segmentation models are used for on-device applications, then segmentation accuracy can be achieved, but the large parameter count and resource requirements make deployment impossible
Solution Approach 1:
The training process is segmented into two distinct stages: a first stage that trains a master model on a master dataset, and a second stage that fine-tunes class-focused models on task-specific datasets. This segmentation allows the system to achieve high segmentation accuracy while keeping the final deployed models compact and suitable for edge devices.
Solution Approach 2:
The system dynamically adjusts model parameters through the two-stage training process. In the first stage, a master model learns general features with a larger parameter set. In the second stage, class-focused models are trained with optimized parameter configurations tailored to specific tasks, reducing the parameter count while maintaining or improving accuracy for targeted classes.
2Measurement precision
If training data is collected from each deployment site to handle domain diversity, then model accuracy improves, but the process becomes non-scalable and delays time to market
Solution Approach 1:
A master dataset encompassing multiple domains is collected and processed in advance during the first stage training. This preliminary action captures diverse agricultural scenarios, crop types, and environmental conditions, enabling the master model to learn robust general features that transfer across different deployment sites without requiring site-specific data collection.
Solution Approach 2:
The master model serves as a universal foundation that can be adapted to multiple different agricultural domains and deployment scenarios. By training on a diverse master dataset that includes various crops, pests, and environmental conditions, the model gains multi-functionality and can be deployed across different farms and regions without retraining from scratch.
3Loss of time
If AutoML tools are run with limited training data, then the process completes faster, but the search fails and generates poor quality, overfitted models
Solution Approach 1:
The first stage training on a comprehensive master dataset performs the preliminary learning of general features and patterns. This preliminary action provides a robust foundation that prevents overfitting when the second stage uses limited task-specific data, as the model has already learned transferable features from the diverse master dataset.
Solution Approach 2:
The master model acts as an intermediary between the general domain knowledge from the master dataset and the specific task requirements. It mediates the transfer of learned features to the class-focused models in the second stage, enabling effective training with limited data while maintaining model quality and preventing overfitting.
Data Source
AI summary
Existing segmentation models have the disadvantage that they cannot be deployed on edge devices as they are not compact and require large space. Embodiments disclosed herein provide a method and system for generating a class focused segmentation model using a two stage model training approach, which enables generation of a compact, task specific model. Using a first stage of training, a Machine Learning (ML) segmentation model that is trained on a master training dataset is generated. Further, in a second stage of training, a fine-tuned task specific segmentation model is generated, which can be used for task specific class segmentation. The fine-tuned task specific segmentation model being task specific and in turn compact, can be used edge device deployment and such applications.