Increasing Precision of Water Sprout Detection based on Mask R-CNN with Data Augmentation

Abstract

This study evaluated the detection performance of four Mask R-CNN models trained in different scenarios. The first two scenarios are trained with a learning rate of 0.01 using data augmentation on the training data. The other two scenarios are trained with a learning rate of 0.001 and the same as previously, using augmentation on training data. These models are trained to detect water sprouts in cacao plants. The original data used are obtained from photographed pictures on the cocoa farm. As much as 150 images, the data is divided into 120 images for training data and 30 images for testing data. In previous studies, the model was trained without performing data augmentation, so that the amount of data trained was less than this study. Data augmentation is implemented to compromise the small amount of data and prevent over-fitting during the model training process. This process uses six augmentation parameters, namely horizontal flip, blur using Gaussian blur, contrast modification using linear contrast, color saturation alteration, cropping the sides of the image randomly by 50 pixels, and rotating the image. The test is carried out by varying the threshold value in the range of 0.6 to 0.9. The results obtained indicate that the model trained with a learning rate of 0.001 with data augmentation can detect objects better than other models with an F1score of 0.966 at a threshold of 0.8. This research will be developed to create a water sprout cutting robot in the future.