Smoke Recognition Algorithm Based On Semantic Segmentation Region Optical Flow Enhancement

The gas produced by straw combustion will cause air pollution and endanger human health.Straw combustion is easy to ignite the surrounding combustibles,trigger deep forest fires and cause huge losses of life and property.Detecting the smoke produced by straw combustion is one of the effective means to monitor straw combustion.Traditional smoke detection methods are difficult to meet the detection requirements in complex scenes,and the existing neural network smoke detection methods are prone to give false alarms to smoke objects such as cloud,fog and lake.Therefore,how to reduce the false positive rate of smoke detection models has become an urgent problem to be solved.In this thesis,the smoke recognition algorithm based on optical flow enhancement in the semantic segmentation region is studied.The fuzzy images that is prone to give false alarms is filtered out according to the fuzziness of the images,and the static smoke background object that is prone to give false alarms is filtered out according to the optical flow information of smoke.The main work of this thesis is as follows:（1）A smoke image detection method based on fuzzy image filtering is proposed.By evaluating the fuzziness of the image and filtering the low-quality fuzzy image with a given threshold,the false positive rate of the model is reduced.At the same time,several mainstream fuzziness evaluation methods are compared experimentally,and the threshold to distinguish fuzzy images from clear images is analyzed.The experimental results show that compared with direct smoke image detection,the accuracy of this method in this thesis is improved by 10%and the recall rate is improved by 4%.（2）A smoke recognition algorithm based on optical flow enhancement in the semantic segmentation region is proposed.The smoke region is located and segmented by using the object detection model and semantic segmentation model.The optical flow characteristics in the region are obtained by optical flow calculation and optical flow statistics.The regional optical flow characteristics are classified by SVM to filter out the false-alarmed static background objects.At the same time,an improved Deep Lab V3+ network is proposed for smoke segmentation.The pyramid feature fusion module is used to fuse the shallow features with the deep features layer by layer,which improves the representation ability of the output features,enables the network to capture more location information in the shallow network,and uses full sampling dilated convolution to suppress the semantic differences of adjacent feature points caused by sparse sampling of dilated convolution.The experimental results show that the improved Deep Lab V3+ semantic segmentation can better locate the segmentation boundary of smoke and get more refined boundary segmentation results.The accuracy of the three test sets is improved by 0.88%,1.42% and 1.51% respectively.The smoke recognition algorithm based on optical flow enhancement in the semantic segmentation region proposed in this thesis can effectively solve the problem of false alarms for static smoke background objects.The smoke detection accuracy（ACC）on the smoke video dataset is improved by 5.75% and the false positive rate（FPR）is reduced by 12.62%.