Abstract: In recent years, deep learning-based image compression methods have surpassed traditional algorithms in multiple benchmarking studies. However, when applied to multi-source datasets containing daily scenes, remote sensing imagery, UAV captures, and noisy visual data, traditional JPEG2000 still outperforms deep learning models, particularly in regions with high-frequency spectral energy or structurally complex features. This paper systematically analyzes over 80 000 images, with statistical verification demonstrating that JPEG2000 achieves superior peak signal-to-noise ratio (PSNR) at equivalent bitrates in some cases. This observation motivates our proposed High -Entropy High-Frequency Vulnerability Hypothesis for deep learning-based compression models. Based on rate-distortion theory, we quantitatively characterize the entropy distribution patterns of latent variables specifically within vulnerable high-frequency spectral bands. To mitigate this limitation, we develop an adaptive frequency-domain classifier for dynamic compression method selection. Experimental results demonstrate that our selector delivers relative PSNR improvements of 0. 43%, 1. 19%, and 5. 87% compared to end-to-end deep compression models on daily scenes, aerial imagery, and noisy image datasets, respectively. The improvement (2. 06 dB) is 2. 1-4. 2 times the just-noticeable difference (JND) of human visual system's threshold.