CHUKWUKA, BEJIH (2026) DESIGN AND IMPLEMENTATION OF AN AUTOMATED BLURRY IDENTITY DOCUMENT IMAGE DETECTION SYSTEM. Other thesis, Godfrey Okoye University, Enugu Nigeria.
![[img]](https://eprints.gouni.edu.ng/style/images/fileicons/text.png) |
Text
CHUKWUKA IBEJIH - FINAL PROJECT REVIEW.pdf Download (1MB) |
Abstract
Identity verification systems in banking, telecommunications, and government registration increasingly require users to upload photographs of identity documents captured on smartphones under uncontrolled conditions. Blur, caused by hand movement during exposure or incorrect focal distance, is one of the most frequent and damaging quality problems affecting such images, as it weakens edge sharpness, destroys fine character detail, and renders documents unreadable for optical character recognition (OCR) and biometric verification. Existing systems accept uploads based on basic technical constraints such as file format, file size, and resolution, none of which detect visual blur, causing blurry images to enter the verification pipeline and produce downstream failures, repeated submissions, and increased manual review workload. This study designed and implemented an Automated Blurry Identity Document Image Detection System capable of assessing image quality at the point of upload and rejecting unusable submissions before verification begins. The system employs a hybrid classical computer vision approach combining Laplacian variance, which measures edge sharpness through second-order derivatives, and Tenengrad gradient magnitude, which evaluates edge strength using Sobel operators. A composite blur score is computed by normalising and fusing both metrics into a single value on a scale of 0 to 100, which is then used to assign one of four quality grades which are Good, Mild, Moderate, or Severe and compared against an administrator-configurable acceptance threshold to produce an acceptance or rejection decision. The system was implemented using a FastAPI backend with an OpenCV-based detection engine and a React-based user interface, with all upload records, scores, grades, and decisions persisted to a PostgreSQL database to support auditing and future threshold optimisation. Evaluation on a 100-image synthetic dataset constructed using Gaussian blur at controlled kernel parameters yielded an overall classification accuracy of 94.00%, a precision of 100.00%, a recall of 88.00%, an F1-score of 93.62%, a false acceptance rate of 0.00%, and an average processing time of 18.26 milliseconds per image, confirming that the system meets real-time performance requirements without requiring GPU hardware. The results demonstrate that the proposed hybrid classical approach outperforms single-metric methods and provides an effective, explainable, and configurable quality gate that reduces verification failures, eliminates unnecessary manual review, and improves user experience in digital identity verification workflows.
| Item Type: | Thesis (Other) |
|---|---|
| Subjects: | Q Science > Q Science (General) |
| Divisions: | Faculty of Computing And Information Technology (FACIT) |
| Depositing User: | Cynthia Ugwuoti |
| Date Deposited: | 06 Aug 2026 09:29 |
| Last Modified: | 06 Aug 2026 09:29 |
| URI: | http://eprints.gouni.edu.ng/id/eprint/6210 |
Actions (login required)
 |
View Item |
