OFDM (orthogonal frequency division multiplexing) is an essential technique in the physical layer of wireless communications, and OFDM system requires rigid orthogonality between subcarriers. However, in practical systems, the imperfection of components like the oscillator and filter would introduce IQ(in-phase and quadrature-phase) imbalance into the system. The IQ imbalance would infect the orthogonality between subcarriers and decrease the system performance. The effect of IQ imbalance was discussed and an IQ imbalance compensation algorithm with the guidance of parallel DNN (deep neural network) was proposed. The deep neural network relies rarely on mathematic models, and the proposed algorithm utilizes this feature to recover the original signal from the received signal in the frequency domain to its original binary sequence of transmitted signal directly. Meanwhile, the prior knowledge that the interference comes from the image aliasing effect was utilized to initialize the model-driven neural network. Simulation results proves that the proposed algorithm can effectively compensate IQ imbalance distortion, and it outperforms traditional LS algorithm based on pilots in both amplitude and phase compensation and proves the superiority of deep learning solutions for issues in the physical layer.