Branch Detection with Deep Learning -Developing Branch Angle Detector- = Branch Detection with Deep Learning -Developing Branch Angle Detector-

다국어 초록 (Multilingual Abstract)

In Japan, decreasing the number of farmers and aging are big problem, so mechanization and automatization of agriculture has been promoted. In fruit growing, measuring tree height and stem diameter is carried out for the purpose of calculating annual growth or deciding proper amount of fertilizer and chemicals. But it consumes a lot of time and labor because it is measured by human power. So, we address automatization of tree mensuration by using image processing. If we can apply it to tree mensuration easily, it will contribute to precision agriculture and improve the quality and quantity of agricultural products.
Our proposing method for tree mensuration consists of three steps. First, detect tree part from pictures of orchard tree. Second, make 3D model from the pictures of tree part. Finally, calculate tree volume from the 3D model. In this study, we focused on the first step, especially branch detection from pictures. However, it is difficult to detect branch directly from picture because it need to distinguish branch from background and other trees. So in this study, we first cope with a method of detecting branch angle with deep learning to make subsequent blanch detection easier. The aim of this study is branch angle detection and classing branch picture by its angle. After classifying pictures by branch angle, it becomes easy to extract the branch from orchard tree image.
We adopted convolutional neural network (CNN) for the classification system that achieved many remarkable results in image recognition field. We developed 8 layer CNN classifier. This was four class (0°, 45°, 90°, 135°) classifier, and achieved a recall of 93.43%. The result shows that machine learning can be used for detecting branch.
To know which part in the image contribute to the judgement, we implemented Grad-CAM that is the visualization tool of CNN system. Through the result of Grad-CAM, we considered that the system didn't pay attention to the branches but may watch the whole of the picture.