is it possible to train a detection model that can detect the skewed/tilt word/license plate?(重新訓練det模型以檢測出歪斜字樣或車牌號碼是可行的嗎?) #14243
t109598032
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
以下圖車牌號碼為例，以en-PPOCRv3-det檢測出來的結果如左上角，這個綠框會趨近直角四邊形，但因為是俯視拍攝，該框內容經仿射變換會變成有點扭曲、歪斜，在某些案例可能導致辨識度不佳，所以想問問是否能重新訓練一個det模型，把標註框改成貼合字體(如左下角)，使得檢測出來的框就是趨近這種形狀，之後接仿射變換能較為正楷，使得辨識更準確?這個想法在目前det網路架構上重新訓練是可行的嗎?
Beta Was this translation helpful? Give feedback.
All reactions