This paper proposes an infrared and visible light fusion method based on gradient residual dense blocks and shuffle attention mechanism to address the problem of insufficient extraction of fine-grained details and difficult utilization of deep features in deep learning-based infrared and visible light image fusion. This method incorporates gradient residual dense blocks and attention shuffle modules into the encoder, which enhances the ability of the autoencoder to extract fine-grained details and deep global features and suppress noise. In comparison experiments with other methods, our method shows good performance in subjective evaluation in terms of detail texture and global level, and it effectively fuses the effective features of infrared and visible light source images. In objective evaluation, our algorithm achieves optimal values in five metrics including standard deviation, peak signal-to-noise ratio, visual information fidelity, QAB/F, and wavelet feature mutual information, which are 76.927 5, 16.775 5, 0.876 7, 0.514 1, and 0.431 3.