Smoke detection is very important in early fire warning. The existing detection algorithms are basically based on deterministic convolutional neural networks. However, deterministic neural networks tend to give very confident prediction results, even though they do not know whether there is a target object in some regions at all. In particular, the smoke edge region is more transparent, making it extremely easy for these areas to be confused with the surrounding environment. Therefore, the detection algorithm cannot make a good judgment on this region, producing a large number of false positives. So, an improved DeepLabV3+ algorithm is proposed. First, the algorithm optimizes DeepLabV3+ based on Bayesian ideas to output non-deterministic feature coding, so as to quantify the measurement of uncertainty in the predicted image and calibrate the learning process of the model. Secondly, feature coding is preprocessed based on the preprocessing idea to reduce the amount of information of irrelevant interfering features, and the feature fusion capability in the DeepLabV3+ network is strengthened to make full use of the multi-scale feature information extracted from the network. Finally, the upsampling operator in the DeepLabV3+ network is optimized as the CARAFE operator to reduce the loss of important information in the upsampling process. The model achieves good performance on the open SMOKE5K dataset, with the MIoU index reaching 92.41%.