一种伪装昆虫图像的前背景自动分割算法——ZDNet  

An automatic foreground-background segmentation algorithm for camouflaged insect images-ZDNet

作　　者：范炬臣 李小林[1] 任昊杰 王荣[3] 张飞萍[3] 黄世国[1,2] FAN Ju-Chen;LI Xiao-Lin;REN Hao-Jie;WANG Rong;ZHANG Fei-Ping;HUANG Shi-Guo(School of Computer and Information,Fujian Agriculture and Forestry University,Fuzhou 350002,China;Key Laboratory of Smart Agriculture and Forestry in Fujian Province University,Fuzhou 350002,China;College of Forestry,Fujian Agriculture and Forestry University,Fuzhou 350002,China)

机构地区：[1]福建农林大学计算机与信息学院,福州350002 [2]智慧农林福建省高校重点实验室,福州350002 [3]福建农林大学林学院,福州350002

出　　处：《昆虫学报》2024年第8期1127-1136,共10页Acta Entomologica Sinica

基　　金：福建省自然科学基金项目(2021 J01125);福建农林大学科技创新专项基金项目(KFb22097XA,KFB23155)。

摘　　要：【目的】昆虫常在色彩、纹理或形态上和背景相似,具有伪装性,识别难度大。本研究旨在探索基于深度学习的伪装昆虫前背景自动分割方法。【方法】将显著目标检测算法(salient object detection algorithm)、大模型图像分割算法(large-scale model-based image segmentation algorithm)以及伪装目标检测算法(camouflaged object detection algorithm)应用于伪装昆虫数据集,该数据集包括10类昆虫共1 900张图片;并进一步针对现有伪装目标检测算法的不足,提出了一种基于DGNet(deep-gradient network)的网络模型改进方法,即ZDNet(zoom-deep gradient network)。在构建该模型时,充分运用图像特征增强、交错图像金字塔、梯度诱导和跳跃式特征融合等技术。利用伪装目标检测公开数据集COD10K与CAMO构建了包含螽斯、蜘蛛等10个目昆虫的图像数据集,结合迁移学习进行网络训练,将经过训练的模型用于分割伪装昆虫。【结果】现有的伪装目标检测模型用于伪装昆虫前背景分割时,其分割性能明显优于显著目标检测模型和大模型分割图像。同时,ZDNet在性能上也明显优于现有的伪装目标检测算法,获得的S度量值、最大F度量值、平均F度量值、最大E度量值、平均E度量值和平均绝对误差(mean absolute error,MAE)分别为0.890, 0.865, 0.824, 0.966, 0.951和0.020。【结论】研究结果证明了ZDNet网络模型能够获得很好的伪装昆虫前背景分割结果,有利于提高昆虫识别的性能,也进一步拓宽了伪装目标检测方法的应用范围。【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 1900 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 exten

关 键 词：昆虫 伪装 目标检测 深度学习 图像分割 深层梯度网络 

分 类 号：Q969[生物学—昆虫学]