Flame Salient Target Detection Of Multidimensional Aggregation Boundary Aware

Fires threaten life safety and bring huge property losses,and seriously damage the ecological environment.With the rapid development of modern industry,the hidden danger of fire is further increased.Therefore,the rapid and accurate detection of flame information using salient flame detection model is very important to protect the safety of life and property and the ecological environment.The traditional saliency detection algorithm is based on graph theory and frequency domain color.These three kinds of saliency detection algorithms extract salient regions to a certain extent,but the algorithm based on frequency domain information is too sensitive to the information of the inner edge of the image,and it is easy to ignore the overall semantic information of salient objects.The algorithm based on graph layout relies on the underlying characteristics to calculate the node weights,and its accuracy largely depends on the robustness of the underlying characteristics.In the algorithm based on color information,the detection accuracy is greatly reduced when there is an object similar to the flame color in the background of the image.The salient target detection algorithms based on deep learning can be divided into locally opened superpixel,local and global depth comparison,cyclic neural network,and full convolutional neural network.However,the accuracy of extraction and edge refinement of flame features is low in multiple complex scenes.A multi-dimensional edge sensing flame salient target detection model is proposed.The edge perception salience target detection model was selected as the basis: the prediction network fusion multi-dimensional aggregation model;Strengthen the ability of acquiring multidimensional flame information;Residual refining structure was improved and self-interactive learning model was introduced to enhance the refining effect of flame edge.The piecewise loss function is set up to solve the problem of class unbalance caused by cross entropy loss in the combined loss function and the spatial inconsistency of the predicted results.The flame image data sets of various complex scenes are constructed to promote the model to comprehensively learn the flame characteristics and improve the accuracy of flame detection.Through the ablation experiment,MAE of feature extraction and edge refinement in the salient target detection model of multi-dimensional aggregated edge sensing flame decreased by 0.0688 and 0.046,F-Measure increased by 0.2237 and 0.2114,and REALX-F increased by 0.2237 and 0.2114,respectively.In the comparative experiments of six general open data sets and eight advanced salient target detection methods,all the indicators are improved to varying degrees,and the performance is excellent.