This work investigates the use of neural networks to achieve two main objectives: compressing image data and simultaneously improving image quality. Leveraging the MNIST dataset, we applied normalization techniques and added noise during training to increase the robustness of the neural network, resulting in effective compression and image enhancement. This dual approach offers significant improvements in both data size reduction and visual quality preservation. This study dealt with the use of neural networks to compress, reconstruct, and improve images by focusing on improving image quality while reducing the size of the data. The research included downloading the Fashion MNIST dataset, normalizing the pixel values while introducing noise for robustness, and training a sequential neural network using Keras technology to encode, reconstruct, and compress the images properly effective. The results showed the performance values of the model in compressing and improving images and evaluating the training progress. We examined the average pixel difference between the original and reconstructed images. The results demonstrate the effectiveness of neural networks in efficiently encoding and reconstructing images while maintaining or improving visual quality by training on noisy and original images, adjusting training periods, and improving Paradigm, This research contributes to the development of image, processing techniques and lays the foundation for further, exploration into enhancing image compression capabilities through, deep learning methodologies.