一种基于Mask-RCNN图像分割的头足类动物角质颚色素沉积量化方法

A Mask-RCNN based quantification method for pigmentation of cephalopod beaks

为自动化精确获取头足类动物的角质颚色素沉积占比,采用Mask-RCNN深度学习的神经网络模型,实现对角质颚及其色素沉积的图像识别和分割,提出了一种基于面积获取的自动化测量角质颚色素沉积占比新方法。首先对角质颚及其色素沉积进行轮廓标注,将所得结果转化成训练集导入到残差网络(Resnet50)中,提取角质颚及其色素沉积的数字特征。基于特征金字塔网络(Feature Pyramid Networks,FPN)将各层特征加以融合;再利用区域候选网络(Region Proposal Network,RPN)对特征加以学习并生成候选框;最后,对候选框进行非极大值抑制(Non-Maximum Suppression,NMS),得到角质颚和色素沉积的候选区域,从而实现了角质颚色素沉积占比的自动化精确获取。结果显示:利用Mask-RCNN土颚分割精度为93.60%,色素沉积精度为92.47%,下颚为91.78%,色素沉积为88.78%。研究表明,Mask-RCNN深度学习网络模型可以较好地测量角质颚及其色素沉积的量化占比,本研究为头足类摄食动物的研究提供参考。

In order to automatically detect the quantitative proportion of pigment deposition in the beaks of cephalopods,the deep learning network model of Mask-RCNN was used to realize the recognition and segmentation of beaks and its pigment deposition,and a new method of quantitatively automatic measurement for pigment deposition proportion in beaks was proposed.First of all,the beaks and their pigmentation were labeled,and the results were converted into a training set and imported into the residual network(Resnet50)to extract the characteristics of the beaks and their pigmentation.Based on the feature pyramid network(FPN),the features of each layer were merged,and then the region proposal network(RPN)was used to learn the features and generate candidate frames.Finally,the candidate frame was subjected to Non-Maximum Suppression(NMS)to obtain the candidate area of beaks and pigmentation,realizing the intelligent detection of the proportion of pigment deposition in beaks.The experiments showed that the segmentation accuracy of the upper beaks of the cephalopods was 93.60%and the pigmentation accuracy 92.47%;the segmentation accuracy of the lower beaks was 91.78%and the pigmentation accuracy 88.78%.The results show that the deep learning network model of Mask-RCNN could get the proportion of pigment deposition in the beaks and its pigment deposition,providing theoretical references for future studies in cephalopod feeding ecology.