All projects
CESI · 5th year · team project
Deep learning: image processing
Three neural networks to automate image processing: classification, denoising and automatic captioning.
CompletedMay 2026 – Jun 2026

A team project by four students (FISA Computer Science, 5th year) for a fictitious company, TouNum, which digitises documents for its clients. The goal: automate its image processing with three deep learning models, built with TensorFlow in a Docker container with GPU access.
Deliverable 1 · Image classification
- Dataset of 41,402 images in five classes (paintings, photos, diagrams, sketches, scanned text), with detection and repair of corrupted files
- CNN with four convolutional blocks (32 to 256 filters, batch normalisation, dropout), trained with Adam, early stopping and learning-rate reduction
- “Photo or not” task: 92.6% accuracy, AUC of 0.974 and F1 of 0.848 on the test set
- Like-for-like comparison with two external architectures (fakv8 and fakv13): 93.7% and 94.7% validation accuracy, versus 92.4% for our CNN


Deliverable 2 · Image denoising
- Starting question: can a network learn to denoise without clean images? Answered with the “Noise-as-Target” strategy (Noise2Noise): the network learns to recover an image from an even noisier version of it (Gaussian noise, σ = 0.04)
- Residual U-Net with skip connections, trained with a combined loss: 0.3 × MAE + 0.7 × (1 − SSIM)
- On the validation set: PSNR of 33.6 dB and SSIM of 0.925
- Comparison with Citronn v1, a compact U-Net by a team member trained in a supervised way: 30.4 dB and 0.872
- Graphical interface to try the denoiser on any image

Deliverable 3 · Automatic image captioning
- COCO 2017 dataset: 118,287 images and about 590,000 captions, 7,000-word vocabulary, 80 / 10 / 10 split made per image to avoid any data leakage
- Pre-trained, frozen InceptionV3 encoder, whose features are extracted once and cached on disk
- Two decoders compared: a GRU with Bahdanau attention and a Transformer
- Measured with the BLEU score, the Transformer beats the RNN (BLEU-4 of 13.7% versus 9.1% with greedy search); beam search (k = 5) clearly improves both models: 33.0% and 24.3%
- Effect of data size: comparison with training on only 6,000 images (BLEU-4 of 5.5% and 6.4%)


The three models are the stages of a single image-processing pipeline: denoise, sort, then describe.