Model Training via Auxiliary Image Data Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models trained using transfer learning are inaccurate when the sample size for the target task is limited, as they rely on insufficient labeled data, especially in image processing where manual labeling is inefficient.
Innovation Solution
The method involves acquiring target and source image data, determining auxiliary data from the source data based on similarity, and fusing it with the target data to train a model using a pre-trained classification network, thereby increasing the dataset without manual labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If transfer learning is used to fine-tune a pre-training model with limited target task samples, then the training process can be completed, but the obtained target model is inaccurate
Solution Approach 1:
The patent merges target task samples with auxiliary samples from source tasks to create a fused dataset. The model is trained on this combined dataset, allowing the limited target samples to be augmented with relevant source samples, thereby improving model accuracy without requiring large amounts of manually labeled target data.
Solution Approach 2:
The patent introduces a sample selection module as an intermediary that automatically selects relevant auxiliary samples from source tasks based on similarity to target samples. This intermediary component bridges the gap between source and target domains, enabling effective knowledge transfer even with limited target samples.
2Reliability
If manual labeling is performed to obtain sufficient labeled samples, then model training accuracy can be improved, but the time and cost increase significantly
Solution Approach 1:
The patent uses automatically selected auxiliary samples from source tasks as copies or substitutes for manually labeled target samples. These copied samples from source domains serve as effective training data for target tasks, eliminating the need for time-consuming manual labeling while maintaining model accuracy.
Solution Approach 2:
The system performs self-service by automatically selecting and filtering auxiliary samples based on similarity metrics without human intervention. The sample selection module autonomously identifies relevant source samples, replacing the manual labeling process and significantly reducing time and labor costs.
3Reliability
If only target task samples are used for training, then the model learns the target task specifically, but the model accuracy remains low due to insufficient data
Solution Approach 1:
The patent makes the training system multi-functional by utilizing both target task samples and source task samples. The model can learn from multiple domains through the fused dataset, improving its ability to generalize and achieve higher accuracy on the target task while efficiently utilizing available data from various sources.
Data Source
AI summary
The present application provides a model training, image processing method, device, storage medium, and program product relating to deep learning technology, which are able to screen auxiliary image data with image data for learning a target task, and further fuse the target image data and the auxiliary image data, so as to train a built and to-be-trained model with the fusion-processed fused image data. This implementation can increase the amount of data for training the model, and the data for training the model is determined is based on the target image data, which is suitable for learning the target task. Therefore, the solution provided by the present application can train an accurate target model even if the amount of target image data is not sufficient.