Download original image
Fig. 2. (Color online) Screenshot of the prediction results of a residual neural network trained. Before training, data were standardized, and then the learning conditions are as follows; (activation) the swish function; (epoch) 20; (optimizer) Adam; (loss) the cross entropy. (a) and (c) the test image was never seen in the training and validation phase. (b) the test image was not seen in the training step before. (d) the test image was seen training and validation phases. (e)-(i) are prediction results of the neural network for (a)-(d). The filter of the UV lamp blocked wavelengths over 400nm, but did not block the visible light range of 380 to 400 nm. The training image contained the spectra of the UV lamp overlapping the spectral regions of PPO and PPO+POPOP.
New Phys.: Sae Mulli 2022;72:582~593