Image Data Generation via Grid Segmentation and Cross-Entropy Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data generation methods for image segmentation models are inadequate, as they rely on extensive training data and are not well-suited for multi-target classification networks, and existing methods either require a batch normalization layer or focus on accuracy over completeness and diversity of generated data.
Innovation Solution
A data generation method that uses grid processing to divide a guide image, randomly assigns class labels, and performs repeated grid training using a cross-entropy loss function to generate data that is usable in the target model, optimizing the input image based on loss and regularization function values until a threshold is reached, without relying on preliminary information from a batch normalization layer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If existing data generation methods are used for image segmentation models, then the model can be trained, but the generated data lacks completeness and diversity
Solution Approach 1:
The guide image is divided into multiple grid areas, and each grid area is independently processed to generate corresponding patch images. This segmentation approach allows systematic coverage of the entire image space, ensuring complete data generation while maintaining the structural integrity and diversity of the original image.
2Ease of manufacture
If data generation methods relying on batch normalization layer are used, then the generation process is simplified, but the method is not suitable for multi-target classification networks
Solution Approach 1:
The data generation method is designed to be universally applicable to various image segmentation models, including multi-target classification networks. By using grid processing and cross-entropy loss function, the method can accommodate different model architectures and classification targets without requiring model-specific adaptations or relying on batch normalization layers.
3Measurement precision
If accuracy is prioritized in data generation, then the generated data is more precise, but the completeness and diversity of the data are reduced
Solution Approach 1:
Different grid areas are processed with appropriate class label assignments based on their specific characteristics and requirements. This local quality approach allows each region to generate data with suitable accuracy while maintaining overall diversity across the entire image, preventing uniformity that would reduce data variety.
4Manufacturing precision
If repeated grid training is performed until threshold is reached, then the data generation is optimized, but the training time increases
Solution Approach 1:
The training process uses the cross-entropy loss function as a feedback mechanism to evaluate the generated data quality. By monitoring the loss value and stopping when it falls below a predefined threshold, the method achieves optimized data generation without unnecessary prolonged training, balancing data quality with training time efficiency.
Data Source
AI summary
A method and an electronic device with image data generating are disclosed. The electronic device includes: one or more processors; and memory storing instructions configured to cause the one or more processors to: input an input image to a target model that performs segmenting on the input image to generate a segmented image whose pixels have respective class labels predicted by the target model, calculate an optimization value for the input image based on the segmented image and based on a class label of a first grid area among a plurality of grid areas of a guide image, and optimize the input image based on the optimization value.