Hello Cyril,
We have got considerably better results using the following approach:
- generated masks based on the actual mesh model,
- estimated image quality for all cameras (with ignore masked regions option enabled),
- disabled cameras with image quality lower than 0.8,
- disabled cameras looking at the model from behind,
- rebuilt texture.
However, some areas with the blurred texture still can be found they are not that obvious as in the original model.
Generally, you may reconsider the image acquisition process: pay attention to focusing on the model's surface (some images are obviously out of focus), and try to avoid taking multiple images from the same camera location - if you are doing that, it is suggested to use calibrated nodal head and group the corresponding cameras to the "camera station".