Research On Image Recognition Of Common Diatoms In Highly Complex Backgrounds Based On Faster R-CNN

Diatom is a kind of single-celled algae plant widely distributed on the earth.It has a variety of forms and mainly lives in the waters around us.In recent years,water pollution in our conutry has become more and more serious.If the concentration of nutrients in the water is too high,algae will multiply and form the red tide that affects the number of aquatic organisms such as fishs.Because of algae are particularly noticeable for water quality changes,accurate recognition of the species and quantities of algae are conducive to water quality monitoring.In addition,there will be a amount of diatoms in the organs of corpses drowned in water.Accurate recognition of the species and quantities of diatoms in organs in forensic examination is an important basis for forensic medicine to judge the cause of death.In the process of recognizing algae in water,sands and stones need to be removed around the algae for obtaining a clean background.Various experts and scholars proposed many different algorithms for recognizing these algae.However,these algorithms are basically for recognizing algae with the simple backgrounds.Unlike the recognition of algae in water,the recognition of diatoms in forensics has the following difficulties:(1)There is a highly complex background interference in the diatom images extracted from the forensic examination,which causes a great interference to the recognition.(2)The manual recognition of diatoms in forensic examination basically relies on the human eye to classify diatom images through a microscope.This method is time-consuming and labor-intensive and is prone to misidentification due to human's subjectivity.Therefore,the automatic recognition on the diatom objects in forensic examination is a challenging issue.Aiming at this issue,we proposes a framework including collect diatom images,establishing image data set,extracting image feature,and detecting diatom objects.The framework uses technical methods to collect diatom images,and then generates multiple diatom data sets through different preprocessing techniques.Finally,the object detection algorithm in deep learning is used to recognize and locate the diatom objects,and finally output the diatom species and location in complex background images.The main contributions of this thesis are as follows:(1)A framework for diatom recognition and localization based on deep learning networks is proposed.In this thesis,through the analysis of the characteristics of diatom images in forensic medical examination,the advantages and disadvantages of existing algae researches are summarized.Combined with the actual application scenarios in this paper,the Faster RCNN algorithm is proposed to recognition and localization diatom objects in diatom images.First,extract the proposal regions,roughly locate the object position,and then extract the features of the proposal regions to recognize and locate the diatoms in the regions accurately.(2)The diatom image datasets with different background interferences are constructed.Most of the existing algorithms for algae recognition employ self-built algae image data sets,which are not disclosed and are not aimed at diatom object recognition in forensic testing.Therefore,after classifying and proofreading the collected diatom images in this paper,the data sets with different background interferences are constructed to verify the performance of the algorithms.(3)Validate the limitations of traditional algorithms for the recognition of diatoms interfered by highly complex backgrounds.The existing algorithms are mainly focused on the recognition of alage in simple backgrounds in water.Few studies are focused on the recognition of diatom images in forensic testing,and no specific algorithm framework are proposed to solve practical problems.Therefore,the datasets with different background interferences are employed to verify the limitations of traditional recognition algorithms based on the traditional machine learning.(4)Optimizing the training strategies of Faster R-CNN and the network models through experiments,improve the performance of diatom recognition and localization.Through comparative tests,the parameter combination with better performance of recognition and localization is selected in this thesis.