Abstract
In this study, the suspected area of downy mildew in onion images was detected. To this end, image processing-based disease area detection method and deep learning-based disease area detection method were used. The evaluation was calculated based on the area, thereby providing criteria for disease detection. 1000 images were collected using the automatic image collection device, and this was sampled at 224×224. As a result of checking the image for 10 days, it was confirmed that the onion leaves affected by downy mildew browned and bent over time. Algorithms were implemented using these features. The image processing-based disease area detection method consists of image noise filtering, HSV conversion, color distance calculation, detection by distance threshold, detection by blob detection, and area rate calculation. As a result of doing this, it was found that browned leaves and surrounding soil were detected together. For deep learning based disease area detection method, VGG16 model was used to transfer learning. And CAM was created by summing and normalizing the feature map and GAP output from the model. As a result of this implementation, unlike the image processing-based method, results similar to the suspected pathogen-determined areas were visually displayed. In conclusion, it was confirmed that the deep learning method was superior to the image processing-based method by automatically outputting the suspected area of downy mildew. In addition, if deep learning is used, it is judged that alerting of downy mildew according to the activation level will be possible.
Figures & Tables
Fig. 1. Downy mildew in onion leaves.
