Style Transfer Model for Training Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for generating training data for intelligent container systems using synthesized sample data often result in inaccuracies due to significant differences between the synthesized and real data, leading to reduced item identification capabilities.
Innovation Solution
A method involving an image style transfer model that minimizes a loss function, comprising an original loss function, background loss function, and foreground loss function, is used to generate training data, ensuring a small difference between the generated and target images, thereby enhancing the accuracy of the training model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthesized sample data is used to form the training data set, then the number of sample data is greatly increased, but the synthesized sample data is quite different from real sample data, resulting in inaccurate training models
Solution Approach 1:
The patent uses style transfer technology to copy the visual characteristics and distribution features from real sample data to synthesized sample data. By transferring the style (color distribution, texture, lighting conditions) from real images to generated images, the synthesized data maintains realistic appearance while preserving the advantages of machine generation, thus resolving the contradiction between quantity and accuracy
Solution Approach 2:
The patent modifies the parameters of synthesized data by applying style transfer transformations that adjust color spaces, contrast, brightness, and other visual parameters to match real sample data characteristics. This parameter adjustment ensures that synthesized data maintains both high quantity and high fidelity to real-world distributions
2Manufacturing precision
If manual photo taking is used to collect sample images, then the sample data quality is high, but the amount of sample data collected is limited due to the large variety and unlimited number of items
Solution Approach 1:
The patent merges the advantages of both manual collection and machine synthesis by combining real sample data (for style reference) with synthesized sample data (for quantity expansion). The style transfer process bridges these two sources, allowing the system to maintain high quality characteristics of manual collection while achieving the large data volumes necessary for comprehensive training
3Speed
If conventional style transfer is used, then the processing speed is fast, but the transferred image has large differences from the target image, reducing training accuracy
Solution Approach 1:
The patent implements a feedback mechanism through the loss function that continuously compares the transferred image with the target image and adjusts the style transfer process accordingly. The loss function calculates differences in multiple dimensions (color, texture, structure) and provides feedback to optimize the transfer, ensuring both speed and high similarity to the target image
Data Source
AI summary
A method, a device and a terminal for generating training data is provided. The method for generating training data includes: obtaining an original image; determining a transferred image based on the image style transfer model and the original image, wherein the image style transfer model is obtained by minimizing a loss function, the loss function is determined by the original loss function the background loss function and the foreground loss function; determining the training data based on the transferred image. The difference between the generated training data and the target image is small, thereby improving the accuracy of the training model obtained by using the training data.