Few-Shot Image Recognition Using Hard Episode Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning models face challenges in achieving high classification accuracy with a small number of samples due to over-fitting issues, particularly in tasks like medical image segmentation where data annotation is costly and rare data is difficult to obtain.
Innovation Solution
A few-shot image recognition method utilizing hard episode training, which involves constructing an image episode with a support and query set, calculating intra-class and inter-class distances, and optimizing model parameters through back-propagation to enhance model stability and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of sample data is used for model training, then the classification accuracy is improved, but the data annotation cost and time consumption increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training a deep neural network model on a large number of general image data before few-shot learning. This pre-training phase prepares the model with general feature extraction capabilities, so that when only a small number of labeled samples are available for specific tasks, the model can already perform effective classification without requiring extensive annotation time.
Solution Approach 2:
The patent segments the training process into distinct phases: pre-training on large-scale general data, then fine-tuning on small-number labeled samples. This segmentation allows the model to first learn general features efficiently, then adapt to specific tasks with minimal annotation, thereby reducing the time cost while maintaining accuracy.
2Loss of time
If a small number of sample data is used for model training, then the data annotation cost is reduced, but the model produces over-fitting problems leading to low accuracy
Solution Approach 1:
The patent uses preliminary action by pre-training the model on large-scale general image data before few-shot learning. This pre-training establishes robust feature extraction capabilities that prevent over-fitting when only a small number of labeled samples are available, thereby maintaining high classification accuracy with minimal annotation time.
Solution Approach 2:
The patent employs copying by using a pre-trained model as a foundation that can be copied and adapted to different few-shot learning tasks. The pre-trained model's learned features are transferred and reused across multiple tasks, allowing the system to achieve high accuracy with very few task-specific labeled samples.
3Productivity
If random inactivation technology is used for few-shot training, then the model can be trained with small samples, but the accuracy of final test results remains low due to over-fitting
Solution Approach 1:
The patent applies preliminary action by pre-training the model on large-scale general data before few-shot learning, establishing robust feature extraction capabilities that prevent over-fitting and maintain high test accuracy even with limited samples.
Solution Approach 2:
The patent uses parameter changes by adjusting the model's training parameters and architecture specifically for few-shot scenarios, including modifying loss functions and optimization strategies to work effectively with small sample sizes while maintaining generalization performance.
Data Source
AI summary
The present application discloses a few-shot image recognition method and apparatus, a device, and a storage medium. The method includes: obtaining to-be-recognized images, and constructing an image episode according to the to-be-recognized image, the image episode including a support set and a query set; inputting the image episode into a pre-trained image recognition model, the image recognition model being a few-shot image recognition model based on hard episode training; and calculating a similarity between an image in the query set and each class in the support set according to the image recognition model, and determining the class of to-be-recognized images in the query set according to the similarity. According to the image recognition method provided by the embodiments of the present application, model training and image recognition can be performed by using fewer image samples, and hard episodes are fused into a training process of a few-shot image recognition model, whereby the few-shot image recognition model can be trained more efficiently and quickly, and the trained model is higher in stability and higher in accuracy of image recognition.