Abstract: Accurate 3D reconstruction from a single RGB image remains a challenging task in computer vision, especially in industrial applications such as quality inspection and virtual assembly, where existing methods still struggle with complex occlusions and multi-object interactions. This paper proposes a novel encoder-decoder framework for single-view 3D reconstruction that integrates a channel attention mechanism with residual convolutional networks to enhance feature extraction and occlusion handling. The encoder incorporates an improved channel attention module to strengthen local feature representation, while the decoder predicts sub-voxel grid displacements along back-projection rays to recover fine-grained structures. Experimental results on the ShapeNet synthetic dataset demonstrate that the proposed method achieves a mIoU of 62.4% for single-object reconstruction at 128 resolution, outperforming CoReNet by 3.3%. For multi-object scenarios, it reaches a mIoU of 48.2%, surpassing CoReNet by 4.3%. In severely occluded scenes (>50% occlusion), the mIoU improves by 4.6%, confirming the robustness of the approach. The proposed method provides a high-precision solution for 3D reconstruction in complex industrial environments.