Deep Learning-Based Image Reconstruction Under Different Sampling Patterns: A Comparative Study of Direct and Unrolled Architectures

Image reconstruction from incomplete measurements is a fundamental problem in signal and image processing, with applications ranging from medical imaging to computational photography. In recent years, deep learning approaches have shown promising performance, particularly when combined with physics-inspired formulations such as deep unrolling. This paper presents a systematic comparative study of classical interpolation and variational reconstruction methods, direct convolutional neural networks (CNNs), and unrolled data-consistency CNN architectures for image reconstruction under different sampling patterns. We consider three representative mask types: structured block masks, nonuniform masks, and random sampling patterns, with sampling ratios ranging from 10% to 50%. Experiments are conducted on the public BSDS500 image dataset, using a fixed grayscale preprocessing pipeline and a reproducible train/validation/test split. Experimental results demonstrate that reconstruction performance strongly depends on the sampling pattern. For random masks, the full unrolled DC-CNN achieves the best quantitative and qualitative performance, reaching a PSNR of 30.98 dB and an SSIM of 0.921 at 50% sampling. In contrast, for structured block and nonuniform masks, TV-based inpainting provides the strongest overall performance, showing that classical model-based reconstruction remains highly competitive when the sampling pattern contains spatially coherent missing regions. A block-size sensitivity analysis further confirms that the difficulty of structured-mask reconstruction is governed by the geometric severity of the missing region. Statistical analysis using paired tests with Holm correction confirms that the main performance differences are significant across the evaluated configurations. Furthermore, we show that a lightweight unrolled model with shared weights and reduced depth achieves a substantially lower parameter count and lower computational cost than the full unrolled architecture, although with reduced accuracy in the most favorable random-sampling cases. These findings provide practical insights into the relationship between sampling strategies and reconstruction performance, offering guidance for the design of efficient and robust learning-based reconstruction systems.