Skip-cycleGAN:一种果园苹果异源图像配准模型  

Skip-cycleGAN:A Heterologous Image Registration Model for Orchard Apple

作　　者：何亚鹏 刘立群[1] HE Ya-peng;LIU Li-qun(School of Information Science and Technology,Gansu Agricultural University,Lanzhou 730070,China)

机构地区：[1]甘肃农业大学信息科学技术学院,甘肃兰州730070

出　　处：《计算机技术与发展》2024年第7期40-47,共8页Computer Technology and Development

基　　金：甘肃省高校教师创新基金项目(2023A-051);甘肃农业大学青年导师基金资助项目(GAU-QDFC-2020-08);甘肃省科技计划资助(20JR5RA032)。

摘　　要：针对有监督的配准模型的性能受限于给定的标签以及循环一致性生成对抗网络训练不稳定,收敛速度较慢,易过拟合,对复杂场景的图像处理效果不佳的问题,基于循环一致性生成对抗网络从3个方面(生成器、鉴别器和损失函数)进行改进,提出一种无监督的异源图像配准模型。生成网络的下采样与上采样之间引入带有特征转换残差层的跳跃连接,可以确保梯度的有效传递,减少前向与反向传播过程中信息损失,实现低级特征和高级特征的结合,从而缓解梯度消失和梯度爆炸,促进神经网络的收敛,有助于网络学习更多的上下文信息。在一个自建果园苹果数据集和两个公共数据集上对模型进行评估,实验得出在改进后的生成器基础上,对于形变比较大的数据集选取70×70 PatchGAN鉴别器更合适,对于形变比较小的数据集选取PixelGAN鉴别器更合适。与8个经典算法进行对比,用6个性能指标进行评估,实验结果表明该模型在异源果园苹果数据集上的综合表现优于对比算法。未来将提升模型对异源图像亮度和对比度的鲁棒性,并进行轻量化模型的工作。Aiming at the problems that the performance of the supervised registration model is limited by the given labels as well as the unstable training of the loop consistency generative adversarial network,which has a slow convergence speed,is prone to overfitting,and is ineffective in image processing for complex scenes,an unsupervised heterologous image alignment model is proposed based on the improvement of loop consistency generative adversarial network from the three aspects of the generator,the discriminator,and the loss function.The introduction of a jump connection with a feature transformation residual layer between the downsampling and upsampling of the generative network ensures the effective transfer of gradients,reduces the loss of information in the process of forward and backward propagation,and achieves the combination of low-level features and high-level features,thus alleviating the gradient vanishing and the gradient explosion,promoting the convergence of the neural network,and helping the network to learn more contextual information.The model is evaluated on a self-built orchard apple dataset and two public datasets,and the experiment concludes that on the basis of the improved generator,it is more appropriate to select the 70×70 PatchGAN discriminator for datasets with relatively large deformation,and the PixelGAN discriminator for datasets with relatively small deformation.Comparing with eight classical algorithms and evaluating with six performance metrics,the experimental results show that the comprehensive performance of the proposed model on the heterologous orchard apple dataset is better than that of the comparison algorithms.Future work will be done to improve the robustness of the model to the brightness and contrast of heterologous images and to lighten the model.

关 键 词：图像配准 异源图像 生成对抗网络 跳跃连接 岭回归损失 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]