基于改进ResNet-UNet的火焰图像分割方法

Research on flame image segmentation method based on improved ResNet-UNet

针对现有火焰图像分割精度低的问题,提出了一种基于改进ResNet-UNet的火焰图像分割网络,实现火焰的精确分割。以UNet神经网络为基础,采用ResNet-34网络作为神经网络特征提取前端,加深了网络层数,保留ResNet-34的独立卷积结构和残差结构,并保证ResNet-34网络和UNet网络的有效融合,从而形成ResNet-UNet神经网络。针对火灾图像中目标区域与背景区域在图像中所占比例差别大而导致的分割类别不平衡问题,将交叉熵损失与Dice损失线性组合并对Dice损失加权。经实验证明,提出的改进ResNet-UNet方法较改进前准确度提高了11%,Hausdorff_95效果更好,网络预测值和真实值更接近,有效提高了火焰的分割性能。

Aiming at the problem of low precision of flame image segmentation,a network based on improved ResNet-UNet was proposed to achieve accurate flame segmentation.Based on UNet neural network,the ResNet-34 network was used as neural network feature to extract front end,the network layers were deepened,its independent convolution structure and residual structure were retained,and the perfect fusion of ResNet-34 and UNet networks were ensured to form a ResNet-UNet neural network.In order to solve the problem of unbalanced segmentation category caused by the large difference in the proportion of target area and background area in the image of fire image,linear combination of cross entropy loss and Dice loss and weighted Dice loss were set up.Experimental results show that the improved ResNet-UNet method improves the accuracy by 11%,the Hausdorff_95 effect is better,and the predicted value obtained by the network is closer to the real value,which effectively improves the segmentation performance of flame.