Foreground retention solutions for low-contrast images
When the subject is close to the background color, the following measures can be taken to ensure the quality of the segmentation:
- pretreatment stage::
- Contrast Enhancement Using CLAHE Algorithm (OpenCV Implementation)
- Histogram equalization of V channels in HSV space
- Manual labeling of a small number of keypoints to guide the algorithm (requires modification of the model input layer)
- Model Application Tips::
- Set the threshold parameter to 0.3 when calling the API (default 0.5)
- Temperature scaling of the confidence output layer (temperature scaling)
- Combined with CRF post-processing to refine the boundaries
- alternative::
- Change to an inference version of the HSV color space (custom loading required)
- Manual correction of masks for particularly important images
Experimental data shows that the segmentation accuracy of the preprocessed white wedding dress sample with white background can be improved from 62% to 89%. it is recommended to fine-tune the domain adaptation for specialized images such as medical/X-ray first.
This answer comes from the articleBEN2: Deep learning model for fast background removal from images, videosThe
Related files download url
You need to log in to download this resource. Go to log in
© Download resources copyright belongs to the author; all resources on this site are from the network, for learning purposes only, please support the original version!
May not be reproduced without permission:AI productivity tools " How to avoid the foreground loss problem when processing low-contrast images with BEN2?
