Research On The Classification Method Of Embryonic Developmental Stage Of Takifugu Fasciatus Based On Deep Learning

The pufferfish has high economic value and occupies an important position in the field of fishery and aquaculture.However,during the breeding process of pufferfish,the embryonic development period needs to be manually observed.Adopting different breeding strategies for embryos in different developmental periods is beneficial to improve the embryo maturation rate and hatching rate.Using deep learning technology to classify and detect embryonic developmental stages can solve the inefficiency problem of manual observation,thereby improving the efficiency of artificial breeding.The dark-patterned pufferfish embryo images collected by the industrial trinocular electron microscope have background light interference,which affects the accuracy of classification and detection,and the detection accuracy of the existing classification and detection network is not high on fewer data sets.Eliminating the background light interference of embryo images and improving the accuracy of the classification detection network is the focus of this paper.This paper proposes(1)combined FMR and CLAHE for image preprocessing of pufferfish embryos and(2)improved YOLOv4 classification of embryonic developmental stages of pufferfish.The main contents of this article are as follows:(1)In order to eliminate the interference of the image background of the dark-striped puffer fish embryo image on the main body of the embryo under the electron microscope,and to highlight the main body of the embryo in the image of the dark-striped puffer fish embryo,this paper proposes a combination of FMR and CLAHE algorithm to enhance the image of the dark-striped puffer fish embryo image.Methods: First,the FMR algorithm was used to enhance the image of the dark-striped pufferfish embryo,and the embryo image with a bright background was obtained without affecting the details of the main body of the embryo;then the image after the FMR algorithm was enhanced by the CLAHE algorithm based on the V channel,and the final image was obtained.Enhance image.(2)Due to the small embryo data set,the classification and recognition accuracy of the original YOLOv4 network is low.In order to improve the classification and detection performance of the YOLOv4 network,this paper introduces the Triplet Attention mechanism on the backbone network of the original YOLOv4 to improve the feature extraction ability;the Focal EIOU Loss function is used.Improve the loss convergence accuracy as a new loss function.The embryo images were divided into five developmental stages,and background light interference was eliminated using the enhancement method in(1)in the image preprocessing stage.Finally,the YOLOv4 network is used to classify and detect the original image and the preprocessed image,and the improved YOLOv4 network is used to classify and detect the preprocessed image,and the performance of the classification and detection is compared and analyzed.Compared with the original image,the average gradient value of the enhanced image is increased by 176.76%,and the average information entropy is increased by9.90%.The overall image enhancement effect is better than other algorithms.The YOLO detection results show that the accuracy of the enhanced image is increased by6.61 percentage points compared with the original image,and the m AP is increased by4.4 percentage points.(2)Compared with the original network,the improved YOLOv4 network has increased the accuracy rate by 10.26 percentage points and m AP by 7.74 percentage points,reducing the classification loss and effectively improving the detection accuracy of target classification.The application provides a reference.