fire-GAN:基于生成对抗网络的火焰图像生成算法

fire-GAN:Flame Image Generation Algorithm Based on Generative Adversarial Network

针对生成对抗网络生成火焰图像质量不高、颜色难以控制的问题,基于HistoGAN算法,提出一种新的火焰生成算法(fire-GAN)。首先,在图像预处理环节添加火焰图像分割,使网络不受背景的干扰,能有效减少生成火焰发生形状变形、颜色失真的情况;其次,提出圆形度损失函数,使网络在训练过程中更加关注火焰轮廓的复杂度;最后,在生成器和判别器中均采取数据增强,使网络在训练过程中保持稳定,避免发生梯度爆炸。经实验测试,fire-GAN生成的火焰与目标火焰的RGB平均误差为2.6%,Fréchet inception distance(FID)为59.23,inception score(IS)为2.81。实验结果表明,fireGAN能生成与目标火焰图像颜色相近、清晰度好、真实性高的火焰图像。

We propose a novel flame generation algorithm,called fire-GAN,based on the HistoGAN algorithm to solve the issues of low quality and complex color control of flame images produced by a generative adversarial network.First,flame image segmentation is introduced in the image preprocessing link to remove background interference from the network,reduce the flame shape distortion and color distortion.Second,the roundness loss function is suggested to increase the focus of the network during training on the intricacy of the flame contour.Finally,data enhancement is implemented in the generator and discriminator to maintain the network stability during training and prevent gradient explosion.The experimental results demonstrate that the average RGB error between the flame generated by fire-GAN and the target flame is 2.6%,the Fréchet inception distance(FID)is 59.23,and the inception score(IS)is 2.81.The outcomes demonstrate the feasibility of the fire-GAN to produce a flame image with color,definition,and authenticity levels quite comparable to the target flame image.