基于条件GAN的复杂动作图像轮廓智能捕捉研究  

Research on complex action image contour intelligent capture based on conditional GAN

作　　者：范蒙蒙 FAN Mengmeng(Jishou University,Jishou 416000,China)

机构地区：[1]吉首大学,湖南吉首416000

出　　处：《现代电子技术》2024年第17期94-97,共4页Modern Electronics Technique

摘　　要：针对复杂动作本身的高动态性和多样性,传统的图像处理方法难以准确捕捉其轮廓,文中研究基于条件GAN的复杂动作图像轮廓智能捕捉方法,精准了解动作执行情况。该方法利用像素覆盖分割模型来分割原始复杂动作图像,获取复杂动作目标图像,将其作为约束条件输入生成器,经过编解码器处理后输出虚假复杂动作图像轮廓生成结果,判别器将生成器输出的虚假轮廓和真实复杂动作图像轮廓作为输入,在损失函数作用下进行真假判别,并采用反向传输的方式对生成器和判别器的参数进行迭代更新,实现最佳复杂动作图像轮廓智能捕捉。结果显示:该方法可以有效捕捉不同类型的复杂动作图像轮廓,通过消融实验证明,改进后的条件GAN模型可显著提升捕捉到的图像轮廓的平滑度和形状保持性,为捕捉复杂动作图像轮廓提供了一种新方法。The complex actions are of high dynamics and diversity,so it is difficult for the traditional image processing methods to capture the contours of the action images accurately.In view of this,a complex action image contour intelligent capture method based on conditional GAN(generative adversarial network)is studied for accurate understanding of the action execution.In this method,a pixel coverage segmentation model is used to segment the original complex action image in order to obtain the complex action object image,which is taken as a constraint condition and input into the generator.After the input image is processed by the encoder-decoder of the generator,the false complex action image contour generation result is outputted.The discriminator takes the false contour outputted by the generator and the true complex action image contour as the inputs,and performs true false discrimination under the action of the loss function.The parameters of the generator and discriminator are updated by reverse transmission iteratively,which aims to capture the optimal complex action image contour intelligently.The results show that the method can capture different types of contours of complex action images effectively.It has been proved by ablation experiments that the improved conditional GAN model can perfect the contour smoothness significantly and shape preservation property of the captured images.To sum up,it provides a new method for capturing the contour of complex action images.

关 键 词：条件GAN 复杂动作图像 轮廓捕捉 生成器 判别器 损失函数 图像分割 

分 类 号：TN957.51-34[电子电信—信号与信息处理] TP183[电子电信—信息与通信工程]