Result and discussion

The quality of the image captured using the camera is affected by the flash, frequency distortion, ambient light and intensity of the camera. The above factors are considered as noise. It can be removed from the image using a deep CNN (Andrey et al. 2017). An enhanced image for the brown spot disease of paddy leaves is shown in Figure 12.12. Image augmentation method artificially creates multiple images for training by combining the different processes such as rotation, translation, shearing, flips, etc. Augmentation is done to increase the number of images in the smaller-size dataset to improve the performance of training. The ImageDataGenerator() is a function from Keras library used to generate augmented images artificially.

The experiment is conducted for different epochs like 1, 10, 20, 30 and 40. Table 12.1shows the training and validation accuracy for different epochs by a four-layer deep CNN on the paddy disease dataset. Figure 12.13shows the comparison (Apoorva Sindoori et al. 2017) of training accuracy with validation accuracy. In the figure, 1, 2, 3, 4 and 5 on the x-axis correspond to 1, 10, 20, 30 and 40 epochs, respectively. The four-layer CNN gives a high training and validation accuracy of 98.893% and 97.847%, respectively. Table 12.2shows the validation accuracy of non-augmented and augmented images for different epochs. In Figure 12.14, 1, 2, 3, 4 and 5 on x-axis correspond to 1, 10, 20, 30 and 40 epochs, respectively. In Table 12.2, the augmented dataset gives a high validation accuracy of 98.43% at 40 epochs than non-augmented dataset.