Vehicle Camera Image Learning via Masking and View Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning techniques using images from vehicle-mounted cameras are limited by the color of the vehicle and the camera's mounting angle, making it difficult to apply results across different vehicles, as the images include partial views and color variations that affect recognition performance.
Innovation Solution
An apparatus and method that masks a fixed area in the image with a pattern image, converts the masked image into multiple views, and performs deep learning using these images to normalize the input for various vehicle cameras, regardless of color or mounting angle.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning is performed using images obtained through a camera mounted on a specific vehicle, then the learning model can be trained with actual vehicle data, but the recognition performance deteriorates when applied to other vehicles with different colors and mounting angles
Solution Approach 1:
The patent applies parameter changes by modifying the image data parameters (color, brightness, mounting angle) through various transformation operations. The learning model is trained with images that have had their color information altered or removed, and with images simulated at different mounting angles, enabling the model to recognize objects regardless of the original vehicle color or camera orientation.
Solution Approach 2:
The patent achieves universality by creating a learning model that can process and recognize objects across different vehicle types, colors, and camera mounting configurations. The masking operation removes vehicle-specific features, and the angle transformation generates multiple viewing perspectives, making the model universally applicable to various vehicles rather than being specialized for a single vehicle configuration.
2Loss of information
If the image includes the color of the specific vehicle and partial image according to mounting angle, then the image represents the actual vehicle view, but the learning result cannot be applied to other vehicles with different colors and views
Solution Approach 1:
The patent extracts and removes the harmful information (vehicle color and specific mounting angle characteristics) from the images through masking operations. By applying masks that cover or remove the vehicle body portions and their color information, the learning model is trained to focus on the target objects rather than the vehicle-specific background, enabling cross-vehicle applicability.
Solution Approach 2:
The patent introduces another dimension by transforming images from a single fixed mounting angle into multiple images with different mounting angles through rotation and perspective transformation. This dimensional expansion allows the learning model to understand objects from various viewpoints, making the results applicable to vehicles with different camera mounting configurations.
3Manufacturing precision
If conventional deep learning uses images with specific vehicle color and mounting angle, then the model learns from actual captured images, but the model exhibits abnormal recognition performance when applied to other vehicles
Solution Approach 1:
The patent applies preliminary action by pre-processing the training images to remove vehicle-specific color and angle information before the learning model is trained. The masking operation is performed in advance to eliminate distracting vehicle features, and angle transformations are applied beforehand to provide diverse viewing perspectives, so the model learns general object recognition patterns rather than vehicle-specific characteristics.
Data Source
AI summary
An apparatus for learning an image of a vehicle camera and a method thereof are provided to apply a result of deep learning to all vehicles regardless of the color of a vehicle and the mounting angle (e.g., yaw, roll and pitch) of a camera. The apparatus includes an image input device that inputs an image photographed by a camera mounted on a vehicle, and a controller that masks a fixed area in the image input from the image input device with a pattern image, converts the masked image into a plurality of images having different views, and performs deep learning by using the masked image and the converted plurality of images.