In the process of electromagnetic interference sources imaging testing using parabolic reflection, the diffraction phenomenon of the system leads to blurred and low resolution images. Interference sources have different resolution capabilities in different areas and frequencies, so it’s difficult to enhance image resolution by using existed super-resolution algorithm. In order to realize the blind recovery of wideband electromagnetic images, a method based on convolutional neural network is proposed. Network training is the process which directly inputs a blurred image and reconstructs a high quality image without assuming any particular blur and noise model. Both experiment and simulation results demonstrate that the convolutional neural network blind recovery method outperforms other blind recovery calculations in different imaging regions of wide frequency band.