Abstract: The latent variable of Invertible Rescaling Network (IRN) uses Gaussian distribution to embed the high-frequency information of the image. Because of the independence and randomness of this approach, the high-frequency information of the image cannot be fully preserved, and the embedding effect is general, which affects the performance of the reconstruction. In order to improve the ability of embedding high frequency information and further reduce the complexity of the model, this paper proposes an improved algorithm based on IRN. Firstly, the dense connection structure and channel attention mechanism are adopted to obtain sufficient feature information and reduce the number of parameters in the feature extraction module. Secondly, the latent variable of the network is designed by high frequency sub-band interpolation in wavelet domain to improve the embedding ability of high frequency information. The results show that compared with IRN, the average Peak Signal to Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM) of the proposed algorithm are improved by 0.380 dB and 0.014 on the four benchmark test sets Set5, Set14, BSD100 and Urban100, the algorithm parameters in this paper are reduced by about 1.64×10⁶ M, the FLOPs are reduced by about 0.43×10⁹ G, and the running time is reduced by 3 ms. It verifies that the reconstruction performance of the proposed algorithm is excellent and the model complexity is low, which has practical value.