Thesis:

Title : A Low-Latency CPU-Based Unified Deblurring and Denoising Model for Improved Iranian License Plate OCR Accuracy

This thesis studies image restoration as a fixed preprocessing stage for Iranian license-plate ‎OCR‎. ‎IRLPUR-Net‎ is a lightweight U-shaped restoration network operating on 224×48 plate crops with 7,766,051 trainable parameters. Three controlled variants were trained with identical architecture and data: A with ‎L1‎ loss, B with ‎L1‎+‎SSIM‎, and C with ‎L1‎+‎SSIM‎ plus a frozen ‎VGG16‎ perceptual loss. The final dataset contained 1,682 unique sharp plate crops split deterministically into 1,346 training, 168 validation, and 168 test images. Six fixed degradation conditions were evaluated per validation/test image. On the independent synthetic test, model C achieved 26.02 dB ‎PSNR‎ and 0.9080 ‎SSIM‎ versus 18.96 dB and 0.6270 for degraded inputs. A separate real-world challenge set of 106 blurry plates, without paired sharp references, was evaluated with a fixed ‎Hezar‎ CRNN ‎OCR‎ model. Exact plate accuracy increased from 3.77% on raw inputs to 30.19% after ‎IRLPUR-Net‎, while ‎CER‎ decreased from 0.4587 to 0.1875. A fixed classical refinement after C further reduced ‎CER‎ to 0.1392 and yielded the highest observed exact accuracy of 33.02%. An official NAFNet-GoPro-width32 checkpoint was evaluated as an external zero-shot transfer baseline; its weak transfer to Iranian plate crops was interpreted as a domain-shift result rather than a general ranking of architectures. CPU benchmarking on an ‎Intel Core i7-6500U‎ using four threads, 20 warm-up runs, and 200 measured runs produced a mean ‎IRLPUR-Net‎ forward latency of 51.02 ms. The measured plate-crop-to-restoration-to-‎OCR‎ sub-pipeline averaged 135.75 ms; upstream detection and disk I/O were excluded. Overall, the results indicate that a domain-specific restoration stage can substantially improve the information available to a fixed ‎OCR‎ system under the evaluated degradations while remaining feasible for CPU inference. Limitations include the size and prior pilot exposure of the real-world set, synthetic degradation modeling, single-seed training, dependence on one ‎OCR‎ evaluator, and an out-of-domain external baseline.

Keywords: Keywords: image restoration; deblurring; denoising; Iranian license plates; ‎OCR‎; CPU inference; character error rate; ‎IRLPUR-Net