Fossil Images Retrieval Based On The Paleontology 3D Models

Fossil image is the information carrier of fossil specimens,which is an important basis for discussion paleontological taxonomy and systematics among paleontologists.Retrieval fossil images accurately and automatically would help paleontologists to establish biological evolution relationships and paleontologists to learn paleontological knowledge.Although using normal computer vision methods could automatically recognize and retrieve fossil images,and can effectively reduce the error rate and subjectivity in the fossil image retrieval process.However,there are still two main problems.The first is that the amount of real fossil image data is insufficient,which leads to the generalization of performance of the model is degraded.The second is that the color difference between the main body of the fossil image and it's background is small,which makes the extracted features contain a lot of noise and would have a lower retrieval accuracy.Therefore,this thesis proposes a fossil image retrieval method based on the three-dimensional model of paleontology to solve both of these problems.Firstly,the paleontology three-dimensional model is used to generate fossil images to supplement the training set.Secondly,the saliency detection method is used to extract the main features of the fossils and combine them with the global features for joint retrieval.The main research contents include.(1)To solve the problem that the insufficient fossil image data leads to the low generalization performance of deep neural networks,this thesis proposes a fossil image data expansion method based on adversarial neural networks.Map the paleontology three-dimensional model into a multi-view two-dimensional image,and mine the relevant information between the mapped image and the fossil image.Through experiments,the fossil images synthesized by the method in this thesis not only have real visual perception,but also the performance of the deep learning network trained by the synthesized images can be greatly improved,indicating that the synthesized images have rich and meaningful fossil image information.(2)To solve the problem that the background of paleontological fossil images is compatible with the subject and would have a large amount of noise if directly extract image feature,this thesis proposes a fossil image saliency detection method based on Transfer Learning.The method of transfer learning is used to get salient map which can detect the saliency of the real fossil image,and obtain the main features of the fossil image.The experimental results show the feasibility and effectiveness of simulated fossil image data and migration learning in the saliency detection of real fossil images.(3)To solve the problem that effective salient features cannot be extracted more complex fossil images,leading to the problem of retrieval errors,this thesis proposes a fossil image retrieval method based on the fusion of salient features and global features.The salient features are extracted by the saliency detection network and the global features are extracted by the classification network,which are merged to jointly express the characteristics of the fossil image.The experimental results show that the method of fusing features can improve the classification accuracy.Using Resnet18 and Resnet50 as the basic network respectively,the classification accuracy is increased by 8% and 11%.In terms of retrieval performance,the average accuracy rate m AP reached 94.4%,and when the error tolerance range was limited,the accuracy rate of top K of the proposed method in this thesis are higher than other retrieval methods.In summary,the proposed method in this thesis enable to achieve identification and retrieval tasks of fossil image,effectively increase the accuracy and efficiency of fossil image retrieval,and can assist paleontologists to do fossils image researching.