Synthetic Image Generation for Anomaly Detection Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The scarcity of anomalous images for training machine learning systems, particularly in surveillance video analysis, limits the effectiveness of anomaly detection due to the prevalence of normal data over abnormal data, making it challenging to obtain diverse and sufficient training data across various weather and lighting conditions.
Innovation Solution
A system and method for synthetic image generation using neural networks, including a light-source-modeling neural network, density-modeling neural network, color-modeling neural network, and occupancy-modeling neural network, which generate images based on weather, time, and pixel coordinates, enabling the creation of both normal and anomalous images under different conditions, thereby addressing the scarcity of anomalous training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning systems are trained using real surveillance images, then the system can learn from authentic data, but the scarcity of anomalous images limits training effectiveness
Solution Approach 1:
The patent creates synthetic copies of normal surveillance images and modifies them to generate anomalous training images. By copying existing normal images and applying transformations (adding objects, changing weather conditions, modifying lighting), the system generates sufficient anomalous training data without requiring actual anomalous events to occur during data collection
Solution Approach 2:
The system changes parameters of existing images to create anomalous variants. By modifying image parameters such as adding detected objects, changing weather conditions (rain, snow, fog), adjusting lighting conditions, and transforming normal scenes into anomalous ones, the system generates diverse training data from limited original images
2Adaptability or versatility
If diverse training data across various weather and lighting conditions is obtained, then model robustness improves, but the complexity of data collection increases
Solution Approach 1:
The system performs preliminary actions by collecting normal surveillance images under various conditions in advance and storing them. During training, these pre-collected normal images are transformed into anomalous images through synthetic modifications, eliminating the need for complex real-time data collection across all conditions
Solution Approach 2:
Instead of collecting diverse anomalous images under various conditions, the system copies normal images and applies synthetic transformations to create anomalous variants. This approach generates unlimited diverse training data from a single normal image through parameter modifications and object additions
Data Source
AI summary
A system and a method are disclosed for synthetic image generation. In some embodiments, the system includes one or more processors; and a memory storing instructions which, when executed by the one or more processors, cause performance of: receiving weather input data; receiving time input data; receiving pixel coordinates; using a light-source-modeling neural network, computing a light source model based on inputs of the weather input data and time input data; and using an image-generating system, generating an image based on the pixel coordinates and the light source model.