Research And Application Of Image Enhancement Algorithm For Takifugu Obscurus Embryos

In recent years,the poor efficiency of takifugu obscurus breeding has caused the motivation of farmers to decrease,and there is an urgent need to further enrich the breeding species of takifugu obscurus in the market.However,the efficiency of artificial breeding and cross breeding of takifugu obscurus is low,and the accuracy of classifying and detecting embryo quality and each period of embryo growth by manual means is not high.It is important to use image processing technology to accurately identify and classify each period of takifugu obscurus embryos and establish suitable breeding environment such as salinity and temperature by analyzing the characteristics of each period to improve the survival rate of artificial breeding and cross breeding,and it is important to adopt automatic method and improve the accuracy of detecting each period of takifugu obscurus embryos.Image enhancement of takifugu obscurus embryos is needed because the acquired images of takifugu obscurus embryos are subject to uneven illumination,halo artifacts,color shifts and noise,resulting in poor accuracy in the detection and recognition of takifugu obscurus embryos.However,the study of applying image enhancement techniques directly to takifugu obscurus embryos is immature and challenging.In this thesis,based on traditional and deep learning image enhancement algorithms,we perform image enhancement of takifugu obscurus embryos with loss of image exposure,halo artifacts and edge information details,and classify and detect embryo quality and each period of embryo growth by target detection algorithms.In this thesis,we use images of takifugu obscurus embryos as the dataset and focus on the following aspects:(1)A study of image enhancement algorithms based on deep learning for the dataset of this thesis: In this thesis,starting from the types of datasets needed for network training,firstly,image enhancement algorithms such as Dn CNN and Real-ESRGAN based on clear and distortion-free(Ground-Truth)image training networks,using the public datasets to train the network to obtain the weight parameters,and testing on the dataset of this thesis to obtain the enhancement effect;secondly,N2 N and Nb2 Nb image enhancement algorithms based on real(Truth)image training networks are trained and tested to obtain enhancement effects using the dataset of this thesis.The results of the analysis of subjective visual and objective evaluation indexes show that only the RealESRNet algorithm can improve the image clarity without reducing the detail information,thus enhancing the image quality,but the effect is not obvious.(2)Enhancement studies based on traditional image enhancement algorithms for the dataset of this thesis: In this thesis,the dark channel prior(DCP)algorithm and its 2inversions as well as derivations are used to enhance the embryonic images.Among them,the improved DCP algorithm derives 4 transmittance rates,and then combines the image pixel reduction and enlargement to finally generate 8 enhancement effects.Through comparative analysis,the A+X combination algorithm combining image brightening and image darkening works best,and the enhanced images are richer in detail information and the features of each period are more obvious and clear.In addition,the method was compared with MSRCR,LRS-CLAHE,DCP and other image enhancement algorithms,and the effectiveness and superiority of the improved DCP image enhancement algorithm was verified by subjective visual and objective evaluation indexes.(3)Study of the dataset of this thesis based on the target detection algorithm.This thesis uses the SSD model to detect the original embryo images with the enhanced images generated by the Real-ESRNet algorithm and the improved DCP algorithm in this thesis,and the experimental results show that the enhanced embryo images based on the RealESRNet algorithm have a better m AP,but the loss function is higher and the localization in detection is not accurate enough;while the enhanced embryo images based on the improved DCP algorithm have the m AP is the highest and the loss function is relatively low,and the detection classification is the best and more accurate.In summary,this thesis applies image processing techniques to the classification of takifugu obscurus embryo images.The experimental results show that the eight image enhancement algorithms based on deep learning mentioned in this thesis are not well applied to the takifugu obscurus embryo images obtained under microscope,but they provide an important reference for future research on takifugu obscurus embryos in the direction of deep learning image enhancement;the traditional DCP-based image enhancement algorithm achieves obvious enhancement effect,can effectively improve the accuracy of embryo detection and classification,and has practical Feasibility.