ED2 dataset

Shape completion remains a fundamental challenge in computer vision and image processing, particularly for tasks involving hand-drawn sketches and occluded objects. Traditional deep learning methods such as Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs) often suffer from high computational costs and poor generalization on sparse, abstract structures. We introduce the Efficiency-Driven Encoder-Decoder (ED$^2$) model, a novel neural architecture designed to achieve state-of-the-art shape reconstruction quality while significantly reducing computational overhead. Unlike conventional methods, ED$^2$ utilizes a compact encoder-decoder framework optimized to minimize structural discrepancies through an adaptive loss function, ensuring high fidelity reconstruction with reduced artifacts. Extensive evaluations on diverse datasets demonstrate ED$^2$’s superior perceptual quality, achieving Structural Similarity Index (SSIM) values of 80–90$\%$ even for inputs with up to 75$\%$ missing information. Furthermore, our approach enhances edge continuity, geometric consistency, and shape plausibility, outperforming state-of-the-art models such as GAN*, SketchGAN*, and U-Net* in both visual realism and computational efficiency. The model’s lightweight design enables real-time deployment in image inpainting, sketch-based design, and augmented reality applications, making it well-suited for resource-constrained systems. By bridging the gap between efficiency and perceptual quality, ED$^2$ sets a new benchmark for robust and scalable shape completion in image processing.