Abstract: 【Aim】Insects often resemble their backgrounds in terms of color, texture, or shape, making them camouflaged and difficult to be identified. This study aims to explore a deep learning-based automatic segmentation method for the foreground and background of camouflaged insects. 【Methods】The salient object detection algorithms, large-scale model-based image segmentation algorithms, and camouflaged object detection algorithms were applied to a dataset of camouflaged insects, which includes 1 900 images across 10 insect classes. To address the shortcomings of existing camouflaged object detection algorithms, an improved network model based on deep-gradient network (DGNet), named zoom-deep gradient network (ZDNet), was proposed. In constructing this model, techniques such as image feature enhancement, staggered image pyramids, gradient induction, and leapfrogging feature fusion were extensively utilized. The insect image dataset, including species from 10 orders like grasshoppers and spiders, was built using public camouflaged object detection datasets COD10K and CAMO. Combined with transfer learning for network training, the trained model was then used for the segmention of camouflaged insects. 【Results】 When the existing camouflaged object detection models were employed for foreground-background segmentation of camouflaged insects, their segmentation performance was markedly superior to those of salient object detection models and large-scale model-based segmentation models. Similarly, ZDNet also exhibited clear superiority in performance over existing camouflaged object detection algorithms, and achieved the S-measure, maximum F-measure, average F-measure, maximum E-measure and average E-measure scores, and the mean absolute error (MAE) of 0.890, 0.865, 0.824, 0.966, 0.951 and 0.020, respectively. 【Conclusion】 The research results demonstrate that the ZDNet network model can achieve excellent foregroundbackground segmentation results for camouflaged insects, contributing to the improvement of insect recognition performance. Furthermore, it extends the application scope of camouflaged object detection methods.

Key words:  Insects, camouflage, object detection, deep learning, image segmentation, deep-gradient network