A Method of Plant Root Image Restoration Based on GAN. Issue 30 (2019)

• Record Type: Journal Article
• Authors: Mi, Jiaqi; Gao, Wanlin; Yang, Si; Hao, Xia; Li, Minzan; Wang, Minjuan; Zheng, Lihua
• Is Part Of: IFAC-PapersOnLine. Volume 52:Issue 30(2019)
• Page Start: 219
• Page End: 224
• Publication Date: 2019
• Subjects: Arabidopsis -- Root system -- Image -- Restoration -- Generative Adversarial Network (GAN); Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2019.12.525
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: Root is one of the most important organs for plants to obtain water and nutrients so that its morphological research is critique for identifying plant growth conditions. Aiming at breakthrough of barriers and obtaining accurate root phenotype data based on the original plant root image, a method of Arabidopsis thaliana root image restoration based on GAN (generative adversarial network) was proposed in this paper. Firstly, a second generation Kinect camera is used to capture the matched data set for training the GAN, which includes high-resolution images of some objects and their matched fuzzy and distort images, and high-resolution images of Arabidopsis' roots and their images in the biogel. Secondly, a GAN with attention mechanism is constructed and trained. The network mainly consists of two parts: the generator and the discriminator with attention mechanism. It is multi-layer convolution network, except that the generator adopts a de-convolution structure to carry out the super-resolution reconstruction. The generator is responsible for converting a fuzzy image into high-resolution image, and the discriminator is used to distinguish whether the inputted image is derived from the prepared dataset or generated by the generator. With the progress of network training, the generator is getting better and better at generating images, the same is true for the effect of the discriminator discriminating the image, that is, the better mapping relationship between the blurred or partially missing image and the high resolution complete image is established. Finally, import the root image of the Arabidopsis planted in the biogel into the trained network and the repaired and restored root image can be obtained. Compared with the original image, the restored one has more accurate details and accordingly more accurate root morphology parameters are computed. The experiment results showed that the proposed method can be used to achieve the super-resolution reconstruction and complete the incomplete or blur Arabidopsis root images.