<p>Image restoration for degradation in complex scenes remains a challenge to obtain high-quality images. To solve the problem of complex degradation of blur and haze in dynamic environments, a method using deep convolutional neural network (CNN) called BHNet based on multi-channel frequency information is proposed for image restoration. Firstly, GoPro-haze dataset for deblurring–dehazing is created for blurry and hazy image restoration. And the characteristic of blurry and hazy color image is analyzed. Then, the BHNet network is constructed by introducing the multi-channel attention mechanism and frequency modulation. The dual-path selective frequency module is incorporated to expand the receptive field, thereby improving global performance. Moreover, the multi-channel implicit frequency-domain attention module is introduced to extract channel-specific information and modulate multi-layer frequency domain features, enhancing the ability to capture deep-level image information. Finally, the experimental results on the public and GoPro-haze dataset demonstrate that the proposed method outperforms previous CNN-based methods, with PSNR and SSIM reaching 28.08 dB and 0.923, respectively. Especially, the proposed method with the training using transfer learning can achieve superior performance and is feasible and effective for blurry and hazy image restoration. Code, dataset and models are publicly available at: <a href="https://github.com/zhehangqiu/BHnet">https://github.com/zhehangqiu/BHnet</a></p>