Quantum imaging based on entangled photon pairs exhibits distinctive advantages in complex environments. Nevertheless, under dynamic inhomogeneous fog scattering conditions, the spatiotemporally random scattering effects of fog fields and inhomogeneous light-field attenuation cause non-stationary degradation of the second-order correlation function, which renders conventional sampling schemes prone to sampling mismatch, degraded imaging resolution and other defects. To address this issue, this paper proposes an inhomogeneous sampling quantum imaging method based on the statistical characteristics of second-order correlation curves. First, full-domain coarse sampling is performed to acquire the second-order correlation curve of each pixel within the imaging field of view, and the peak-to-background ratio (PBR) is extracted to construct the corresponding spatial distribution matrix. Second, a two-dimensional sliding window is introduced and combined with the Chebyshev inequality for cyclic iteration to eliminate potential target outliers induced by scattering noise, thereby estimating the statistical mean and variance of the PBR for local pure background pixels. Afterwards, dynamic detection thresholds for each window are adaptively solved according to the Neyman-Pearson (NP) decision criterion to generate a spatial binary mask matrix distinguishing targets from backgrounds. Finally, inhomogeneous fine sampling is implemented over target regions to accumulate effective signals, and the reconstructed image is formed by fusing the fine-sampled target data with the coarse-sampled background results. Simulations and experimental results demonstrate that the proposed method can effectively suppress non-stationary background noise and improve imaging quality under dense fog scattering conditions. Compared with traditional quantum imaging (TQI) and two-step adaptive sampling optimized (TASO) quantum imaging, the proposed method achieves peak signal-to-noise ratio (PSNR) improvements of 11.49 dB and 9.24 dB, and root mean square error (RMSE) reductions of 0.31 and 0.096, respectively, while the structural similarity index measure (SSIM) remains consistently above 0.6 These results verify the outstanding anti-scattering interference capability and environmental robustness of the proposed method.