In recent years, with the improvement of computer performance and the possibility of large-volume data preservation, the field of Computational Ethology, which uses computers and information science to analyze animal behavior, has always been active. Object detection methods using deep learning have become the mainstream in object detection and tracking in images and animations. In this deep learning, a large amount of training data is required. Moreover, in the object detection method, as the correct data, it is necessary to specify the domain of the object from within the image, and label them separately. In previous studies, the accuracy decreased when dense bee images were detected. Therefore, the aim of this study is to improve the bee detection accuracy of dense states using datasets made in previous studies. Firstly, part of the bee area is classified according to the image cut from the animation at hand. Based on the image of this classification, VGG16 is used to detect the bee area. In addition, the detection area is taken as annotation data to learn how to detect the bee by SSD, an object detection method of deep learning.