融合注意力机制的GAN病虫害图像超分辨率重建
Super-resolution reconstruction of GAN pest and disease images fused with attention mechanisms
费加杰 1杨毅 1曾晏林 1蔺瑶 1贺壹婷 1黎强 1张圣笛1
作者信息
- 1. 云南农业大学大数据学院,昆明 650500
- 折叠
摘要
收集咖啡和柑橘病虫害样本图片,利用TensorFlow深度学习框架,在原始SRGAN(Super-resolution generative adversarial networks)的超分辨率重建网络里加入了注意力模块,对重建图像视觉质量和峰值信噪比(PSNR)、结构化相似性(SSIM)指标进行分析.结果表明,设计的模型和原始SRGAN模型对比之后峰值信噪比提高了2.23,结构相似性提高了7%.在细节纹理方面可以获得更好的视觉效果,重建后的图像识别准确率提高了约4.42个百分点.因此,设计的模型可以对小样本性质的植物病虫害样本进行扩充.
Abstract
The sample pictures of coffee and citrus pests and diseases were collected,and an attention module was added to the super-resolution reconstruction network of the original SRGAN by using TensorFlow deep learning framework.The visual quality,peak sig-nal-to-noise ratio and structured similarity index of the reconstructed image were analyzed.The results showed that the peak signal-to-noise ratio of the designed model was improved by 2.23,and the structural similarity was enhanced by 7%,after comparing with the original SRGAN mode.Better visuals could be obtained in terms of detail texture,and the accuracy of the reconstructed image clas-sification was improved by about 4.42 percentage points.Therefore,the model designed could be used for the expansion of samples of plant pests and diseases with small sample properties.
关键词
超分辨率重建/注意力机制/病虫害/峰值信噪比(PSNR)/结构化相似性(SSIM)Key words
super-resolution reconstruction/attention mechanism/pests and diseases/peak signal-to-noise ratio(PSNR)/structur-al similarity(SSIM)引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据