Developed a lightweight FSRCNN-based super-resolution model using Memory-Aware Automatic Differentiation and Conjugate Gradient optimisation to reduce memory usage during training. Achieved PSNR of 67 dB and SSIM of 0.77 across benchmark datasets.