Xeon processor

The hierarchical CycleGAN incorporating the reconstruction GC loss was trained by solving the mini-max problem, given by $$\begin{array}{c}\hfill {\widehat{G}}_{\mathit{Xp}},{\widehat{G}}_{\mathit{DRR}}=\text{arg}\underset{G_{\mathit{Xp}},G_{\mathit{DRR}}}{min}\underset{D_{\mathit{Xp}},D_{\mathit{DRR}}}{max}{\mathcal{L}}_{\mathit{total}}(G_{\mathit{Xp}},G_{\mathit{DRR}},D_{\mathit{Xp}},D_{\mathit{DRR}})\end{array}$$ using Adam³² with a learning rate of 0.0002 for the first 100 epochs and linearly decreasing to 0 for the later 100 epochs.
The generator was trained hierarchically with nine residual blocks, and the discriminator was $142\times 142$ PatchGAN³³ . The balance parameters were experimentally determined to be ${\lambda }_{\mathit{cyc}}=10$ and ${\lambda }_{\mathit{GC}}=1.0$ .
The average training time for the conventional and proposed methods was approximately 8.0 and 9.7 h, respectively (2.7 and 7.0 h for the training of global ( $G_1$ ) and local ( $G_2$ ) mappings, respectively), on a workstation equipped with an Intel Xeon processor (2.30 GHz, 4 cores) with NVIDIA Titan RTX (24 GB memory). The average computation time for the inference for one case was about 2.5 s, excluding the file access.

The training and testing of the model were performed on a CentOS 7 Linux workstation equipped with one Intel Xeon Processor CPU (96 GB RAM), accelerated by one Nvidia GeForce GTX 1080 Ti GPU (11GB Memory). The model is implemented in the Keras 2.2.4 deep learning open-source framework with the TensorFlow-GPU 1.8.0 backend using Python 3.6. The detection model on each color space took an average of 25 hours for training.
For creating the ground-truth dataset VGG Image Annotator (VIA) [41 ], a standalone software was used for the manual annotation of the blight and leaf patches in the image. It allows a rectangular- and polygonal-shaped area to be annotated, which is useful for training Mask R-CNN.