A Deep Learning project that implements Content-Based Image Retrieval (CBIR) functionality using a Convolutional Autoencoder on the CIFAR-10 dataset. The system learns compressed feature representations (embeddings) of images to clean noisy inputs and retrieve visually similar images efficiently.

• Autoencoder Architecture: A convolutional neural network that compresses images into a lower-dimensional latent space.
• Image Denoising: Capable of removing noise from images as part of the reconstruction process.
• Similarity Search: Retrieves the most similar images from the training set for a given query image using Euclidean distance in the latent space.
• Performance Metrics: Evaluates retrieval performance using Mean Average Precision (mAP).

CBIR/
├── src/                # Source code modules
│   ├── data_loader.py  # CIFAR-10 loading and preprocessing
│   ├── model.py        # Autoencoder architecture definition
│   ├── train.py        # Training procedure
│   ├── evaluate.py     # Evaluation metrics and scoring
│   └── visualization.py# Image plotting and saving
├── models/             # Directory for saved models
├── data/               # Data files (not tracked by git)
├── output_images/      # Generated results (denoised/retrieved images)
├── main.py             # CLI Entry point
└── requirements.txt    # Python dependencies

1. Clone the repository (if applicable) or download the source.
2. Install dependencies:

   pip install -r requirements.txt

The project is controlled via the main.py script.

Train the autoencoder from scratch.

python main.py train --epochs 20 --batch_size 128

This will save the model to models/autoencoder.h5.

Calculate the Mean Average Precision score.

python main.py evaluate --n_test_samples 1000

Perform a search for a specific test image.

python main.py demo --index 1 --n_samples 10

Results will be saved in output_images/.

Visualize the denoising capability of the autoencoder.

python main.py denoise

The model learns to map 32x32 RGB images to a robust feature space.

• Denoising: The model can effectively remove Gaussian noise from inputs.
• Retrieval: Queries return semantically similar images based on visual features (color, shape, texture) rather than just raw pixel matching.

This project is open-source and available for educational purposes.