Multi-Stage Convolutional Autoencoder with Adaptive Quantization and Secure Hash-Based Encryption for Optimized Image Compression and Retrieval

This paper presents a Multi-Stage Convolutional Autoencoder Adaptive Quantization with Secure Hash-Based Encryption (MSCAE-AQ-SHE) for enhanced image compression and retrieval. The model is trained on the CIFAR-10 dataset, which contains 60,000 RGB images across 10 classes. Preprocessing scales pixel values to the [0, 1] range and resizes images to 32 × 32 pixels. Convolutional autoencoders learn compressed latent representations, thus able to save storage while retaining visual quality. The training utilizes the Adam optimizer (learning rate = 0.001), while learning rate scheduling and early stopping are employed to prevent the model from overfitting. Experimental results demonstrate that the proposed system can compress images by up to 85% while maintaining a PSNR of 30 dB or higher, indicating minimal loss in image quality. Adaptive format selection dynamically chooses among JPEG, PNG, or WebP to store the images, balancing size versus quality. A two-layer cryptographic technique based on SHA-256 and MD5 hashing algorithms helps maintain integrity and prevent unauthorised access, thereby enhancing the system. User interaction on Streamlit and retrieval, secure picture storage, and metadata management utilising SQLite are also significant areas of concern. By combining deep learning, adaptive quantisation, and cryptographic security, this provides a highly efficient and secure solution for contemporary image compression and retrieval applications.