Neural Network Training Set Enhancement via Pre-Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The quality of training data and annotations for deep neural networks is often limited by human limitations, particularly in challenging environments such as adverse lighting, extreme weather, and particulate matter, which affects the accuracy and robustness of the neural network's ability to recognize objects and perform tasks.
Innovation Solution
A system that enhances raw digital items, such as images, through techniques like contrast adjustment, de-noising, and edge detection, creating an enhanced training set that allows for more accurate tagging by human operators, which is then used to train a neural network, enabling it to recognize objects without requiring real-time image enhancement during production.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real-time image enhancement is applied during production, then object recognition accuracy in challenging conditions is improved, but processing time and computational resources are increased
Solution Approach 1:
The patent applies image enhancement techniques during the training phase rather than during production. Enhanced images are created in advance and used to train the neural network, so the network learns to recognize objects in enhanced conditions without requiring real-time enhancement during deployment. This transfers the enhancement operation from production to training time.
2Reliability
If more diverse training data with various enhancements is used, then neural network robustness and accuracy are improved, but data preparation complexity and cost are increased
Solution Approach 1:
The patent applies various image enhancement transformations that modify image parameters such as brightness, contrast, saturation, sharpness, and other visual properties. By systematically varying these parameters during training data preparation, the neural network learns to recognize objects under diverse conditions, improving robustness without requiring manually curated diverse datasets.
Solution Approach 2:
The patent creates enhanced versions of existing training images by applying enhancement algorithms. Instead of requiring original diverse images to be captured in various conditions, the system generates synthetic enhanced copies of available images, effectively multiplying the training data diversity from a limited base set.
3Measurement precision
If manual annotation is performed on enhanced images, then annotation quality and accuracy are improved, but annotation time is increased
Solution Approach 1:
The patent performs image enhancement before the annotation process. By pre-enhancing images to improve visibility and clarity of objects, annotators can more easily and accurately identify and label objects in challenging images. This preliminary enhancement step reduces the time and effort required for accurate annotation.
Data Source
AI summary
Embodiments for training a deep neural network by applying tags to enhanced versions of a raw data set are disclosed. In one embodiment, the raw data set may comprise, for example, images, and the method may comprise receiving the raw data set of images, applying an enhancement to the images, deriving tags based on the enhanced images while maintaining a correlation to the original raw images, and using the raw images and the tags to train a deep neural network.