Dual-channel feature disentanglement for identity-invariant facial expression recognition