| Phytoplankton diversity monitoring is an important part of biological assessment of water quality.At present,the main identification method is to determine the species of phytoplankton by manual microscopy,and the experimenter needs professional training,which is time-consuming and labor-intensive.Therefore,the development of automatic identification technology of phytoplankton microscopic images is of great value to the aquatic ecosystem.The fluorescence-assisted bright-field microscopic image recognition technology attempts to accurately identify phytoplankton cells at the species level by fusing phytoplankton microscopic fluorescence images and bright-field images.However,it is difficult to obtain clear microscopic images of phytoplankton in practical applications,and there are some problems such as dislocation of phytoplankton cells in synchronous measurement of fluorescence-bright field images.To address the problems,the accurate recognition method of fluorescence-bright field microscopic image was studied in this paper,and the main research work is as follows.(1)The fusion method of microscopic multi-focus image of phytoplankton cells was studied.Measuring the image of phytoplankton cells requires a high power microscope.Due to the limited depth of field of the microscope,the phytoplankton cells located outside the depth of field have the problem of blurring caused by defocus diffusion.Multiple focal plane microscopic images were acquired,and the focus region and the out-of-focus region were detected by Laplace energy and guide filtering and an S channel of HSV color space.Then,the fusion method of algae microscopic multi-focus image was studied and compared with wavelet transform.Laplace pyramid and pulse coupled neural network fusion methods.The experimental results show that the edge information retention,spatial frequency and average gradient of Anabaena sp.,Scenedesmus sp.and Pediastrum sp.are 0.3529,8.9654,0.0055.0.3778,7.0058,0.0023 and 0.2940,1.5445,0.0005 respectively,which are better than the compared methods.The method effectively realizes the fusion of the microscopic multi-focus images of phytoplankton cells,and provides a method for obtaining the full field depth microscopic images of phytoplankton.(2)The registration method of microscopic bright field and fluorescence synchronous measurement images of phytoplankton cells was studied.The rigid transformation was used as the spatial transformation model of fluorescence image and bright field image.The normalized mutual information of the S channel binary image of bright field HSV color space and the binary image of fluorescence was used as the similarity between the bright field image and the fluorescence image.Then,particle swarm algorithm was used to roughly register the low-frequency components of the five-level wavelet decomposition.The translation and rotation angle of the preliminary registration were taken as the initial values,and the low frequency components of the wavelet three-level decomposition was further registered by using the Powell algorithm.Scenedesmus sp.,Selenastrum carpricornutum,and Nostoc sp.were used as experimental objects.The similarity and registration methods are compared and analyzed respectively.The experimental results show that the mismatch rates are 0.0%,9.4%,and 6.5%respectively,and the average registration time are 10.43s,27.98s,and 17.02s respectively,and the normalized mutual information after registration are 0.673,0.495,and 0.631 respectively.The method proposed in this paper has obvious advantages in registration accuracy and running time,which lays a foundation for phytoplankton identification by fusing fluorescence and bright-field images.(3)In order to solve the problem that the labeling work of phytoplankton instance segmentation method based on MASK-RCNN model needs to manually draw the cell contour in LabelMe and other labeling tools,an automatic labeling method based on fluorescence image is proposed.Firstly,morphological processing such as binarization was applied to the microscopic fluorescence image of algal cells,and then each connected region could be detected and the contours of the algal cell could be automatically drawn in the binary image,so that the fluorescence image is converted into masks required for training Mask-RCNN.Seven algae species were selected as the experimental objects,and the Mask-RCNN network was trained with the data set labeled by LabelMe manually and the dataset constructed by fluorescence automatic annotation,respectively.Finally,the experimental results were compared and analyzed.The experimental results show that the average precision,average intersection over union,and average recall of the test set are 93.28%,98.35%,and 92.02%,respectively,when the Mask-RCNN is trained with the automatic labeling method.The experimental results corresponding to manual labeling are 92.22%,98.25%,and 90.96%,respectively.The experimental results of the automatic annotation method in this paper are almost consistent with those of manual annotation,which verifies the feasibility of the method and provides an effective solution to the problem of the high cost of manual annotation.(4)A method for recognition of phytoplankton fluorescence-bright field images based on convolutional neural network was studied.First,the phytoplankton cells were accurately segmented by fluorescent-assisted bright field,and the phytoplankton microscopic image data sets of ten species,such as Pediastrum sp.and Melosira sp.,were constructed.Then phytoplankton classification and identification methods based on AlexNet,VGG16,MobileNet,ResNet50 and GoogleNet were studied.In order to solve the problem that the training of convolutional neural network requires a large number of phytoplankton image data,transfer learning was used to train the model.The public dataset ImageNet was first used for pre-training,and then the phytoplankton microscopic images were used to fine-tune the last few layers of the model.The experimental results show that the recognition accuracy of five convolutional neural networks for phytoplankton microscopic images is 95.72%,97.04%,96.71%,99.01%and 97.70%respectively,and the recognition accuracy of ResNet50 is the highest.(5)The previous research results were applied to the self-developed phytoplankton monitoring instrument,which realized the automatic scanning,automatic focusing,fluorescence-assisted bright-field segmentation of phytoplankton cells and the identification and classification at the species level.This paper took Nushan Lake as an example to carry out outfield experiments.The actual water samples of Nushan Lake and Chaohu Lake were measured and verified,and 12 species belonging to 10 genera of 3 phyla in Nushan Lake and 20 species belonging to 18 genera of 4 phyla in Chaohu Lake were measured,and the average identification accuracy was 96.88%and 91.94%,respectively,which was consistent with the manual microscopic examination results.In this paper,the automatic and accurate identification of phytoplankton microscopic image was realized by fluorescence-assisted bright-field image.The results provide a new technique for accurate identification of phytoplankton at the species level. |
PDF Full Text Request