关键词: 蝴蝶, 显著性目标检测, 深度学习, 图像分割, 自动分割, F³Net

Abstract: 【Aim】 It is difficult to segment the foreground-background of butterfly images with complex backgrounds. This study aims to explore an automatic foreground-background segmentation method using deep learning based salient object detection method. 【Methods】 The F³Net salient object detection algorithm was trained by using the DUTS-TR dataset to obtain the foreground-background prediction model. Then the model was applied to the dataset of butterfly images with complex background to implement automatic foreground-background segmentation of images. To further improve the accuracy of automatic segmentation, transfer learning was utilized by keeping ResNet backbone unchanged and retraining network through cross feature module, cascade decoders and pixel sensitive loss module to optimize model parameters, and then the better automatic segmentation model was obtained. Meanwhile, other five salient object detection algorithms based on deep learning were also applied to automatic segmentation and compared with F³Net on performance. 【Results】 With all the algorithms good butterfly foreground-background segmentation results were obtained. Among these algorithms, F³Net was the better algorithm, and the algorithm got the values of 0.940, 0.945, 0.938, 0.024, 0.929, 0.978 and 0.909 for the seven indexes S-measure, E-measure, F-measure, mean absolute error (MAE), precision, recall and average IoU, respectively. Transfer learning further improved the values of the above indexes of F³Net, which were 0.961, 0.964, 0.963, 0.013, 0.965, 0.967 and 0.938, respectively. 【Conclusion】 The experimental results showed that F³Net with transfer learning is the best segmentation algorithm. The method developed in this study can be applied to automatic segmentation of insect images taken in field investigations and extends the application range of salient object detection methods.

Key words: Butterfly, salient object detection, deep learning, image segmentation, automatic segmentation, F³Net