EVALUASI KUALITAS VISUAL DAN EFISIENSI KOMPRESI: JPEG VS CONVOLUTIONAL AUTOENCODER (CAE) PADA DATASET DIV2K

Convolutional autoencoder; Deep learning; Image compression; JPEG; PSNR.

Abstrak

Image compression is a fundamental technique in digital image processing aimed at reducing storage space requirements and transmission bandwidth without excessively sacrificing visual quality. As the volume of image data grows dramatically, driven by the expansion of social media platforms, intelligent surveillance systems, and medical imaging applications, the need for efficient and adaptive compression methods becomes increasingly critical. This study presents a comparative evaluation between JPEG, the industry standard based on Discrete Cosine Transform (DCT), and Convolutional AutoEncoder (CAE) based on deep learning, using the DIV2K benchmark dataset (799 high-resolution images, 80:20 train-test split). The proposed CAE architecture employs a 16×16 spatial convolutional bottleneck in two capacity variants: ch=64 (BPP=2.0, CR=12×) and ch=128 (BPP=4.0, CR=6×). Evaluation was conducted using five quantitative metrics: PSNR, SSIM, MSE, BPP, and Compression Ratio, reported as mean ± standard deviation (N=100). Experimental results demonstrate that JPEG still outperforms CAE quantitatively, achieving a peak PSNR of 37.104 ± 3.469 dB at Q=90, while the best CAE variant reaches only 20.703 ± 4.453 dB. This performance gap is attributed to limited training data, simplified architecture, and the absence of explicit rate-distortion optimization. This study provides a rigorous, reproducible quantitative evaluation framework as a baseline for future deep learning-based image compression research.