It is proposed to address the accuracy degradation in deep learning-based pig weight estimation caused by image occlusion in group-housed environments through an image restoration methodology. The method involves restoring occluded images prior to weight prediction, thereby improving estimation accuracy. A specialized dataset comprising 29,104 occlusion-free RGB-D images was constructed from operational pig farms, with artificial occlusions introduced to simulate real-world conditions. The final curated dataset contains 22,451 training images, 2,623 validation images, and 2,765 test images. Three image restoration models are compared under a denoising autoencoder (DAE) framework—CNN, Transformer, and hybrid CNN-Transformer. Results demonstrate that the hybrid model achieves superior performance in both RGB and depth modalities, showing Peak Signal-to-Noise Ratio (PSNR) values of 24.01 dB (RGB) and 20.62 dB (depth), with corresponding Structural Similarity Index Measure (SSIM) values of 0.93 (RGB) and 0.89 (depth), outperforming both standalone CNN and Transformer models. To validate the improvement effect of image restoration for weight estimation, it was implemented to employ an RGB-D fused CNN_224 model for weight estimation experiment. The experimental results show Mean Absolute Percentage Error (MAPE) values of 5.98% for original images, 14.82% for occluded images, and 8.25% for restored images. It was confirmed through practical validation with real occluded images that the proposed methodology effectively improves weight estimation accuracy in group-housing images, providing technical support for high-precision real-time monitoring systems.