CNN-WOA: A Metaheuristically Optimized Convolutional Autoencoder for Practical Learned Image Compression on Commodity GPUs

The explosion of visual data in mobile, cloud, and autonomous systems is increasing the demand for image compression techniques that deliver high perceptual quality while remaining practical on commonly available hardware. Although recent advancements in learned image compression (LIC) have focused on increasingly complex transforms and entropy models such as delta-function entropy modeling for machine image coding, multi-reference entropy models, transformer-based context modeling, wavelet-domain codecs, diffusion-based probabilistic entropy models, and dictionary-based entropy models there has been surprisingly little attention to systematic hyperparameter optimization through metaheuristics or to deploying these methods under realistic training-time constraints on standard GPUs. This paper introduces a convolutional autoencoder-based image compression system that optimizes its key hyperparameters using the Whale Optimization Algorithm (WOA). The model is designed for efficient training on typical single-GPU setups, completing epochs in just a few hours using mixed-precision training and reasonable batch sizes. When evaluated on ۱۲۸×۱۲۸ RGB images from the COCO ۲۰۱۷ dataset, the system achieves a PSNR of A-Y dB, an SSIM of _,۹, and compression ratios of:), while keeping training loss below. ۱°. The proposed CNN-WOA framework differs from existing entropy-model-centric LIC methods yet can be layered on top of them to automatically adjust learning rate, latent dimensionality, and regularization strength. Through a structured comparison with a recent delta-entropy-based codec, we show that our method significantly improves hyperparameter automation, system transparency, and training-time reliability, while maintaining competitive reconstruction quality within realistic computational limits.

Keywords:

learned image compression , convolutional autoencoder , Whale Optimization Algorithm , hyperparameter optimization , entropy modeling