A Quantitative Method Of Pigmentation In Beak Of Cephalopods Based On Deep Learning

Beak is the main feeding organ of cephalopods.Due to the change of feeding habits,the pigmentation on the beak is constantly deposited during their growth.Previous studies have shown that pigmentation can reflect the ecological information of cephalopods,such as day age,carcass length,weight and so on.At present,there are two main research methods of pigmentation in the beak.One is to qualitatively divide it into eight levels based on whether there is pigmentation in the specific part of the beak.But the method is greatly disturbed by human factors.The other is to measuring groups of shape parameters of beak.Then give the correlation between parameter and pigmentation grade.This method needs multiple calculations due to many measurement parameters which the correlation between parameters and grade may conflict.However,the image recognition technology based on deep learning theory has developed rapidly in many fields.Mask-RCNN model represent image recognition algorithm can achieve good results in target detection.Therefore,in this thesis,we present a method to automatically measure and calculate the percentage of pigmentation in beak by using the Mask-RCNN deep learning theoretical framework.Further,we give the quantitative relationship between beak pigmentation and shape parameters.The main contents are as follows:(1)The obtained beak samples were cleaned and saved,and then photographed with a microscope.We get the raw data of the original images.However,the number of Enoploteuthis chunii is not enough to meet the needs of a large amount of data for deep learning.Therefore,in this thesis,we firstly uses migration learning to pre-trains the Mask-RCNN deep learning model by using the beak samples of cuttlefish in Indian Ocean and northwest pacific.Then,we fine-tunes the model parameters of the MaskRCNN method based on the data of the Enoploteuthis chunii.To this end,first of all,the beaks and their pigmentation are labeled,and the results are converted into a training set and imported into the residual network to extract the characteristics of the beaks and their pigmentation.Then based on the feature pyramid network,the features of each layer are merged,and then region proposal network is used to learn the features and generate candidate frames.Finally,the candidate frame is subjected to Non-Maximum Suppression to obtain the candidate area of beaks and pigmentation,and realize the intelligent detection of the proportion of pigment deposition in beaks.The deep learning network model of Mask-RCNN can get the proportion of pigment deposition in the beaks and its pigment deposition,which provides a theoretical basis for the study of cephalopod feeding ecology.(2)Although the Mask-RCNN model has high accuracy in segmenting beak images,it requires a long time for training and prediction.Further,rostrum section of the beak has a shark shape,which induces a poor accuracy of segmentation.Therefore,in this thesis,we use VGG16 network to replace Res Net,and change the full connection layer with the local receptive field.It will reduce the model parameters to decrease the time of training and prediction.To deal with the poor accuracy of segmentation in the rostrum section of the beak,we propose the Canny algorithm to fit the mask branch of the Mask-RCNN model.By these methods,we reduced the time of training and prediction for the improved Mask-RCNN model proposed in this thesis.The segmentation accuracy of upper and lower jaw and its pigment deposition is improved by about 1%.Then,based on the ratio data of the pigment deposition area obtained by the above method,we can obtain the correlation between the pigment deposition and appearance parameters by using the correlation coefficient analysis and grey correlation degree analysis,in order to find the key parameter of the shape to describe the pigment deposition.It is found that the upper rostrum length(URL)is the most closely related to the proportion of pigmentation in the morphological parameters of beak,and the lower hood length(LHL)is the most closely related to the proportion of pigmentation.It follows from the above experimental results that the research method proposed in the paper shows excellent performance in the percentage of beak pigmentation.Based on the correlation analysis,we give the key shape parameters most closely related to the beak pigmentation,which can be used to provide new research ideas to study the day-age,carcass length,weight and other physiological parameters.