Deep Learning Dataset Augmentation via Dynamic Detection Area Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Related-art emotion recognition technologies face performance degradation when using different face detectors for training and inference due to mismatched facial area detection, leading to overall performance issues in emotion recognition systems.
Innovation Solution
A method for augmenting a training dataset by randomly transforming detected areas in images, using a detection area generator to update and refine coordinates, ensuring consistent performance across different detection scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a pre-set facial detection area is used for training, then training efficiency is improved, but the model becomes sensitive to detector performance degradation during inference
Solution Approach 1:
The patent applies dynamics by transforming the static, pre-set detection area into a dynamic, randomly generated detection area during training. Instead of using a fixed facial region, the system randomly selects and transforms detection areas from the training images, making the training process adaptive and robust to variations in detector performance during inference.
Solution Approach 2:
The patent changes the parameter of detection area coordinates by randomly generating new coordinates based on the original detection area. This parameter transformation creates varied training samples where the model learns to handle different detection scenarios, improving its resilience to detector degradation without sacrificing training efficiency.
2Stability of the object's composition
If the same detection area is used for both training and inference, then consistency is maintained, but the model cannot handle detector performance degradation
Solution Approach 1:
The patent applies preliminary action by pre-generating and storing multiple transformed detection areas during the training phase. These pre-computed varied detection areas are then used during inference to handle detector performance degradation, allowing the model to maintain accuracy even when the actual detection differs from the training detection areas.
Solution Approach 2:
The system transitions from a static detection area approach to a dynamic one where detection areas are randomly transformed and generated during training. This dynamic approach prepares the model to handle variability in detection results, maintaining both consistency in training and robustness in inference.
3Reliability
If detection area coordinates are randomly transformed during training, then model resilience to detector degradation is improved, but training data complexity increases
Solution Approach 1:
The patent uses copying by creating synthetic training data samples that replicate real detection scenarios. Instead of working with original complex detection data, the system copies and transforms detection area coordinates to create simplified, varied training samples that capture the essence of detection variability without the full complexity of original data.
Solution Approach 2:
The system manages data complexity by applying parameter transformations to detection area coordinates. By randomly generating new coordinates based on simple mathematical operations rather than complex data processing, the system achieves data variety and model resilience while keeping the training data structure relatively simple and manageable.
Data Source
AI summary
There is provided a training dataset augmentation method and system for training a deep learning model. A training dataset augmentation method according to an embodiment configures a training dataset with image data from which a specific area is detected, and a label, and adds a new training dataset by transforming the detected area in the configured training dataset. Accordingly, by augmenting a training dataset of a deep learning model, which analyzes areas detected from an image by a detector and performs inference, through random transformation of a detection area, the deep learning model may be made to be resistant to performance degradation of the detector.