Smoke opacity is an important indicator for evaluating emissions from stationary pollution sources, but conventional visual inspection is still affected by observer subjectivity, illumination conditions, and background complexity. In this study, we propose a vision-based framework for the automatic estimation of smoke opacity levels from visible-light images. The proposed method consists of four sequential stages: smoke-region segmentation, smoke-type classification, local patch-based opacity estimation, and final decision making. First, a U-shaped convolutional network (U-Net) model is used to extract smoke regions and suppress irrelevant background interference. Next, a support vector machine classifies the segmented smoke as black smoke or white smoke. In accordance with the classification result, sampled local patches are processed by the corresponding convolutional neural network model to predict patch-level opacity classes. Finally, for each sampled patch, a top-k weighted strategy selects the k highest-scoring opacity classes and combines them using normalized score-based weights to obtain a continuous opacity estimate; the patch-level estimates are then averaged to derive the final image-level opacity value. To evaluate the proposed framework, experiments were conducted on a collected dataset with physically measured opacity labels. The results show that the classification-based model achieved average testing accuracies of 90.16% for black smoke and 92.55% for white smoke. When the proposed top-k weighted method was used to estimate continuous opacity values, the accuracies reached 90.6% for black smoke and 96.0% for white smoke within an error tolerance of ±10%. These results indicate that the proposed framework provides an effective and practical approach for automated smoke-opacity estimation under real-world environmental conditions.