I have VHR imagery, 5 cm pixel, and I want to create a DEM but this image has high trees and when I take my model I cant take a real elevation because DEM is not the height of the terrain.
Could I eliminate this effect?
There's not really enough information in your question.
If you have many overlapping images you can use something like Pix4D to generate a 3D model of some form and an orthoimage using structure from motion algorithms. But, unlike LiDAR, this will not be classified.
As far as I know, you can't automatically remove trees and buildings from a 3D model generated from this kind of algorithm. Plus, imagery does not see through trees as LiDAR can (to some extent). So, you'll probably end up with a model that has lots of canopy and very little ground, depending on what the tree cover is like.
Your question does suggest you already have a DEM, and so perhaps you could try classifying the dataset to identify areas that are ground, and then exclude other areas and interpolate a new model based on that, but it's not something that is standard.