ML Scene Detection Model for Portrait Background Replacement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional chroma key technology is limited in replacing backgrounds in portrait photographs, as it requires specific saturated colors and struggles with patterns and subjects wearing similar-colored clothing, making it difficult to accurately distinguish subjects from backgrounds.
Innovation Solution
A system and method using a machine learning-based photographic scene detection model to automatically detect and classify backgrounds, floors, and props of various colors and designs, allowing for replacement without relying on saturated colors or patterns, by training the model with diverse sample images under different conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If chroma key technology is used to replace backgrounds, then background replacement can be achieved with saturated colors, but it fails when subjects wear similar-colored clothing or when backgrounds have patterns
Solution Approach 1:
The patent transforms the background from a simple color-based solution to a structured pattern-based solution. By embedding detectable patterns (such as grid patterns, radial patterns, or other geometric designs) into the background, the system enables reliable detection and replacement regardless of the subject's clothing color or the background's base color. This parameter change from color saturation to pattern structure resolves the contradiction between reliability and adaptability.
2Difficulty of detecting and measuring
If conventional detection methods are used, then simple color-based backgrounds can be detected, but complex designs and patterns cannot be accurately distinguished
Solution Approach 1:
The patent introduces an intermediary detectable pattern as a mediator between the background and the detection system. This embedded pattern acts as a visual cue that facilitates accurate detection and segmentation. The pattern serves as an intermediate element that the detection algorithm can reliably identify and use to define scene boundaries, thereby improving measurement precision without increasing detection difficulty.
3Measurement precision
If machine learning models are trained with diverse sample images, then detection accuracy improves for various conditions, but training time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-training the machine learning model with a comprehensive dataset that includes diverse lighting conditions, subject characteristics, and background patterns before deployment. This preliminary training ensures the model is already adapted to various scenarios, reducing the need for extensive fine-tuning or retraining when encountering different拍摄 conditions in production, thereby mitigating the time loss associated with training.
Data Source
AI summary
A method of photographing a subject includes storing a library of photographic scene designs in a computer memory, training a photographic scene detection model by a computer processing device using machine learning from sample portrait images comprising known photographic scenes defined in the library of photographic scene designs, capturing a production portrait photograph, using a digital camera, of a subject in a photographic scene that is defined by a photographic scene design in the library of photographic scene designs, automatically detecting the photographic scene in the production portrait photograph using the photographic scene detection model operating on one or more computer processors, and processing the production portrait photograph by an image processing system to personalize the photographic scene detected in the production portrait photograph.