Hi Hashem, I would say the main point is that Accord.NET has a higher level of abstraction than the libraries you mentioned, and unlike them it is a unified framework where you can treat images, text, audio, under similar interfaces. And those interfaces also happen to depend on very fundamental .NET types such as jagged and multidimensional matrices, so even if you need to interop with other libraries such as OpenCV or CNTK you do not have to rewrite your application from scratch to use them. You can select things to use from Accord.NET and from OpenCV for example (that's part of the reason for the "accord" name).
Also, since the main goal is to offer an unified experience, Accord.NET can eventually be extended to use those libraries. For example, the .Neuro project will soon be updated to use TensorFlow/CNTK under the hood, but hopefully with a simpler API.
If I had to pick an example I would say that Accord.NET is like scikit-learn for Python, except that it also includes more fundamental libraries such as for mathematics and image processing if you need.