Under complex scattering conditions, obtaining clear target images from captured speckles is highly challenging. To address this, a deep learning framework based on the pix2pix generative adversarial network is proposed for speckle image reconstruction. The experiment includes five groups of focus conditions to capture speckle images with varying correlations. The first four groups are used for network training, with 20% of the data selected as the validation set, while the fifth group is excluded from training to verify the generalization ability of the framework. The results show that even for weakly correlated speckle images with mutual information values less than 1, as well as speckle images captured under unseen scattering conditions, the network can achieve effective reconstruction, with SSIM and PSNR reaching 0.81 and 18.5, respectively. This method demonstrates the network's strong generalization ability under complex scattering conditions, providing new insights for practical imaging applications.