Deep learning techniques for automatic butterfly segmentation in ecological images

Abstract

Automatic identification of butterfly species has attracted more and more attention due to the increasing demand for the accuracy and timeliness of butterfly species identification. Since the butterfly images we captured are usually ecological images, which not only have butterflies but also contain many irrelevant objects, such as leaves, flowers and other complex backgrounds. Therefore, segmenting butterflies from their ecological images is an issue that needs to be addressed prior to the tasks of identification and the segmentation quality directly affects the identification effect. However, the huge differences in butterflies, and the complexity of the natural environment make it very challenging to accurately segment butterflies from ecological images. Deep learning based methods are more promising for butterfly ecological image segmentation than traditional methods because they have powerful feature learning and representation ability. However, butterfly segmentation is still challenging when complex background interference occurs in images. To address this issue, we propose a dilated encoder network to capture more high-level features and get high-resolution output, which is both lightweight and accurate for automatic butterfly ecological image segmentation. In addition, we adopt the dice coefficient loss function to better balance the butterfly and non-butterfly regions. Experimental results on the public Leeds Butterfly dataset demonstrate that our method outperforms the state-of-the-art deep learning based image segmentation approaches.