Digital holographic images are prone to distortion during the compression process, which degrades the quality of the reconstructed images and increases the difficulty of speckle noise removal. Moreover, existing speckle denoising methods require training different networks for varying degrees of compression distortion, limiting their flexibility in practical applications. In response to these issues, an unsupervised speckle noise removal method is proposed for digital holographic reconstructions towards compression distortion. Based on the quality factor of the compression process, a compression perception attention module is designed to dynamically adjust the network weights, achieving adaptive control over the denoising process. Secondly, a speckle noise correlation convolution is used in conjunction with a blind spot network architecture to prevent detail loss due to sampling density constraints. Finally, a nonlinear attention module is introduced to enhance the interaction of global information, allowing the network to more accurately capture speckle noise. Experimental results demonstrate that the proposed method effectively removes speckle noise while preserving more detailed information. Compared to existing methods, the proposed method achieves better results in both objective quality assessment and subjective visual perception.