In view of the problem that the scene information is not be expressed accurately in the visible intensity image, a polarization image fusion algorithm based on residual dense blocks and attention mechanism is proposed in this paper. The proposed algorithm network consists of an encoder, a fusion module and a decoder. In the encoder, a residual dense block is constructed to preserve more feature information and enhance network stability. In the fusion module, channel attention mechanisms are embedded in the intensity feature map extraction network, while spatial attention mechanisms are embedded in the polarization degree feature map extraction network. The Sobel operator is employed to extract gradient information from shallow feature maps that make more feature information retained. In the decoder, the feature maps in the encoder are skip-connected to corresponding convolution layers in the decoder to retain more feature information. Experimental results demonstrate that the fused images obtained by the proposed algorithm not only achieve the best values in multiple objective evaluation metrics, but also have better visual effects and more conform to the human visual perception.