Aiming at the problems of low brightness, low contrast and poor visual effect in images collected in low-light environment, a low-light image enhancement algorithm based on cross-level adaptive feature fusion is proposed. Firstly, a network frontend is built by combining hierarchical sampling and large receptive field convolution to generate multi-scale features of large-area receptive fields, so that shallow information mining can be fully carried out. Secondly, a multi-head transposed attention module embedded in the middle of the network is introduced, the cross-covariance between channels is calculated to generate attention maps, and global context information associations are implicitly established. Thirdly, a joint loss function is constructed to correct the convergence direction of the model, assist the model optimized from the perspective of contrast and structure, and improve the robustness of the algorithm. Relevant experiments are carried out on the LOL and LOLv2 datasets. The experimental results show that the proposed algorithm outperforms most advanced algorithms in terms of objective indicators such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Subjectively, the image brightness is natural and the noise is low, and artifacts are effectively suppressed.