CONTINUOUS OBSERVATION METHOD OF FLOATING OBJECTS AT RIVER SURFACE USING U-NET

Abstract

 Using U-Net, which is one of the semantic segmentation methods, we constructed a model for distinguishing floating objects (scum) at river surface from fixed point camera images, and evaluated the performance based on RMSE, missed rate, and misrecognition rate regarding water surface coverage. The discriminant model was able to reduce the missed rate sufficiently by learning a small amount of image data (about 100 images). However, it was necessary to train a larger amount of data to reduce the false recognition rate. Although the discriminant model was applicable to the images taken at multiple sites, the discriminant accuracy was greatly affected by the angle of view of the camera and the quality of the learning data. By applying transfer learning, it was possible to construct a model with relatively high discrimination accuracy by learning less image data than normal learning.