Occluded Image Detection via GAN Feature Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data augmentation methods are inadequate for occlusion detection tasks, particularly in finger occlusion scenarios, as they fail to effectively generate training data and require extensive manpower and computational resources, leading to low detection accuracy and high costs.
Innovation Solution
A method using a data feature augmentation network and an occluded-image detection model, trained with a generative adversarial network, to generate occlusion image features from a small number of annotated occluded images, improving detection accuracy and reducing resource requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual collection of occluded images is used, then detection accuracy can be improved, but the required manpower and time increase significantly
Solution Approach 1:
The patent uses image synthesis technology to generate virtual occluded images that copy the characteristics of real occluded images. The synthesis model creates fake occluded images from non-occluded images by predicting and filling in occluded regions, thereby obtaining training data without manual collection. This resolves the contradiction by providing sufficient training data (improving detection accuracy) while eliminating the time-consuming manual collection process.
Solution Approach 2:
The patent replaces the mechanical manual collection process with an automated image synthesis system. Instead of requiring human annotators to manually collect and label occluded images, the system uses deep learning models to automatically generate synthetic occluded images through pixel-level prediction and image synthesis, thereby reducing time cost while maintaining data quality for accurate detection.
2Quantity of substance
If existing data augmentation methods are used, then training data quantity can be increased, but the distortion in occluded areas reduces effectiveness
Solution Approach 1:
The patent employs a feedback mechanism where the synthesis model continuously refines its predictions by comparing generated occluded regions with the actual occluded image characteristics. The model uses loss functions to measure the difference between predicted and actual occluded areas, iteratively adjusting the synthesis parameters to reduce distortion and improve data effectiveness while maintaining increased training data quantity.
Solution Approach 2:
The patent changes the fundamental parameters of data generation from traditional geometric transformations to pixel-level synthesis with occlusion awareness. The synthesis model uses occlusion probability maps and conditional generation parameters to create realistic occluded regions that preserve semantic information, thereby improving data effectiveness while increasing training data quantity through virtual generation.
3Measurement precision
If comprehensive occlusion detection is implemented, then detection accuracy improves, but computational resources and training costs increase
Solution Approach 1:
The patent segments the occlusion detection task into multiple independent modules: occlusion probability map generation, image synthesis, feature extraction, and classification. Each module processes specific aspects of occlusion detection separately, allowing efficient computation. The segmentation enables the system to achieve comprehensive detection accuracy by combining results from specialized modules without requiring computationally intensive unified processing.
Solution Approach 2:
The patent applies partial action by focusing computational resources on the most critical aspects of occlusion detection rather than processing every pixel and region equally. The synthesis model generates occluded images only in regions where occlusion is detected, and the classification network focuses on distinguishing occluded from non-occluded regions. This partial processing approach maintains high detection accuracy while significantly reducing overall computational resource consumption.
Data Source
Figure 1~2C
Figure 3A~4
Figure 5~6
AI summary
The present disclosure relates to a method for detecting an occluded image, a device and a medium. The method includes: after an image is captured by a camera, obtaining (101) the image as an image to be detected; inputting (102) the image to be detected into a trained occluded-image detection model, the occluded-image detection model is trained based on original occluded images and non-occluded images by using a trained data feature augmentation network; determining (103) whether the image to be detected is an occluded image based on an occlusion characteristic about an image in the occluded-image detection model; and outputting (104) an image detection result.