TadabburLens
Computer vision / Mobile AI
Recognize a letter, right on the phone.
An Android application that classifies Hijaiyah letters with a MobileNetV2 model running through TensorFlow Lite.
Developer / Project Author
Academic computer vision project · 2025
.Pq6I7F5O_Z9wT1O.webp)
The problem &
my contribution
The scientific project connects a 29-class Arabic letter classifier with a mobile experience: capture an image, review a prediction, and correct the result when needed.
Dataset preparation, transfer learning, model evaluation, and Android model integration.
- Python
- TensorFlow
- MobileNetV2
- TFLite
- Kotlin
- Android
Start with a defined classification task
The project uses 45,308 images across 29 Hijaiyah classes. MobileNetV2 is fine-tuned through transfer learning.
Run inference on the device
TensorFlow Lite places the classifier inside the Kotlin application, with capture, prediction, manual correction, a letter gallery, and history.
Evaluate the model and the experience
Validation accuracy and the confusion matrix describe model behavior. User Acceptance Testing provides a separate view of the mobile flow.
What the evaluation shows
Validation accuracy measures the classifier on the project’s validation set. User Acceptance Testing evaluates the mobile experience separately.
- training images
- 45,308
dataset_v5 · 29 classes
- validation accuracy
- 98.41%
- UAT score
- 88.57%
Inside the application
Original screenshots from the TadabburLens experience. Open any image for a closer look.
.Pq6I7F5O_Z2fbL0C.webp)
.png.kl4XLLbr_1tCPra.webp)
View the remaining 7 application screens
.n-ZWwXwX_Z1osrCw.webp)
.png.CZIgT8IG_2cjBw6.webp)
.CY-X_R7b_bw15j.webp)
).png.DoEzdgfH_Ys2PS.webp)
Evaluation evidence
The original figures and testing records behind the reported results.
Original classifier evaluation across 29 classes. This is a model result, not a measure of the entire application.
Training and validation curves from the project, read alongside the reported validation accuracy.
The source portfolio records 45,308 images across 29 classes. This is not the size of a separate test set.
Original user acceptance results, reported separately from model validation accuracy.