Research On Multi Class Microfossils Target Detection Method Based On SSD Network

The Cambrian in China is rich in paleontological micro-fossils,and many biological species which are completely different from modern organisms have appeared.They are crucial for the evolution of early life on earth and the Cambrian life explosion.However,due to the small size of individual microfossils,the traditional microfossil sorting work is done by manually observing and selecting them one by one under a microscope,which is inefficient.In recent years,image recognition technology has been continuously developed and applied in various fields.In this context,the use of image recognition technology to study microfossils will undoubtedly provide great convenience.Aiming at the identification of microfossils in different scenarios,this thesis proposes an SVM fossil image recognition algorithm based on SIFT features and an SSD network detection algorithm fused with attention mechanism to classify and detect the fossil images collected manually and by machines.In the SVM fossil image recognition algorithm based on SIFT features,firstly,the watershed algorithm is used to segment a single sample from the artificially collected mixed fossil image,and then its SIFT features are extracted.In order to strengthen the expression ability of extracted features,we clustered the SIFT features of all samples to form a "visual dictionary".According to the number of the feature words of each sample and the total number of words,each feature word is weighted to strengthen the expression of features.Finally,the weighted features are used to train the multi-classification support vector machine to classify and recognize the micro-fossil images of each category.And the result is good on the micro-fossil images photographed by human.In the SSD network detection algorithm that integrates the attention mechanism,a microfossil data set is constructed using images collected by machines with adhesion and overlap,and the types and locations of the micro-fossils in the data set are marked.Then analyze the difference between our data set and the public data set.According to the features of the scale and aspect ratio of the micro-fossil images,the original SSD network is improved,and the attention mechanism is integrated for the first two feature layers of the improved SSD network.As a result,the accuracy of network model detection is improved,network training time is shortened,and the amount of network parameters is reduced.As a result,the accuracy of network model detection is improved,network training time is shortened,and the amount of network parameters is reduced.The final network model obtained 93.45% m AP on the micro-fossil data set,which is 5.54% higher than the original SSD network model,and the training time is 16.2 hours,which is an increase of 12.3% compared to the original network model.The amount of parameters is 83.2M,which is 10.8% less than the original network model.And this model can detect micro-fossil images at 39 FPS speed.