Chronic atrophic gastritis is a precancerous lesion that can progress to gastric cancer. Existing deep-learning methods often downsample high-resolution images during diagnosis, leading to a loss of texture details and a reduced ability to identify pathological features of chronic atrophic gastritis. To address this limitation, we propose a novel approach for intelligently identifying chronic atrophic gastritis directly from high-resolution endoscopic images. Specifically, we design a two-stage network that emulates the diagnostic workflow of clinical experts, which first localizes the lesion area to downscale the high-resolution image and then applies a patch-based magnification technique to detect finer pathological features. On a real clinical cases dataset, our method achieves a diagnostic accuracy of 94.08%, with specificity and precision of 96.86% and 96.28%, respectively. These performance metrics outperform many state-of-the-art methods and closely match those of senior endoscopists, providing clinicians with a more objective and reliable tool for lesion assessment.