A Deep Learning-Guided Multi-Scale DCT Compression Framework for Edge Devices

Comparison of image transmission quality through acoustic channel demodulation. From left to right: original images, results from baseline traditional demodulator, proposed Neural Network demodulator, and proposed Neural Network demodulator with fine-tuning. The fine-tuned model demonstrates superior Bit Error Rate and SSIM.

An adaptive image compression method that combines multi-scale DCTs with CNNs. AdaptDCT dynamically chooses DCT block sizes based on image content—larger blocks for uniform areas, smaller for detailed regions. A CNN predicts compression quality (SSIM) and file size to optimize parameters efficiently. Tested on Google's Coral Edge TPU, the method outperforms standard JPEG with better quality at smaller file sizes. In review at IEEE Access.