Few-Shot Detection Of Common Diatoms In Complex Background

Diatoms are an important and widely distributed unicellular algae phytoplankton.The detection of the species and content of diatoms can provide a reliable basis for the diagnosis of drowning in forensic examinations.The earliest diatom detection research was based on traditional laboratory inspection methods,using the cell structure characteristics of diatoms in organ tissues to withstand extreme environments such as strong acids,and using chemical digestion and other methods to remove other substances to retain diatom cells.With the rapid development of computer technology and the improvement of computing power,scientists have begun to turn their research direction to computer-aided diatom detection methods.The shape,texture and other features of diatom cells are extracted manually,and the classifiers such as SVM(Support Vector Machine)and decision tree are used to classify and identify them.As deep neural networks shine in the field of computer vision,scientists have used the powerful feature extraction capabilities of deep neural networks to achieve automatic feature extraction and automatic recognition,and obtain high detection accuracy.However,in deep neural network-based diatom detection methods,high accuracy often requires a large amount of supervised information,that is,labeled diatom image samples.The acquisition process of diatom images is complicated,and it is often difficult to acquire diatom images due to the limitation of the nature of drowning cases.Moreover,the labeling of diatom images also requires a lot of time and manpower.Therefore,for the drowning diagnosis scene in forensic examination,it is of great significance to train a diatom detection model with high detection accuracy using only a small number of diatom image samples,which will help speed up the diagnosis of cases.In order to realize that the object detection model has the characteristics of small number of training samples and high detection accuracy,so as to reduce the time and labor cost of preprocessing operations such as diatom image acquisition and labeling,this thesis proposes common diatom image detection for small samples and complex backgrounds.These include the following aspects:(1)The research background and significance of diatom detection and few-shot object detection were analyzed,and the research status of diatom detection and few-shot object detection was discussed.Diatom detection can assist forensics in drowning diagnosis.By analyzing the development process of diatom detection methods,in view of the difficulty of obtaining diatom images,the research direction of diatom detection,which pursues a small number of training samples and high detection accuracy,is obtained.Few-shot object detection is based on this direction.By analyzing the existing work on few-shot object detection,it can be seen that it is feasible to apply few-shot object detection to diatom detection tasks.Therefore,this thesis proposes the few-shot detection of common diatoms in complex backgrounds.(2)To address the problem of inadequate feature extraction of diatom images by few-shot diatom detection models under the condition of a small number of diatom training samples,the NSNMHSA(Non-Size Normalization Multi-Head Self-Attention)is proposed to improve the model feature extraction backbone network Res Net-101,which is called NSN-Bo TNet-101.First,the multi-head self-attention mechanism is introduced based on Res Net-101 to construct a feature extraction backbone network Bo TNet-101 that can fully utilize the local and global information of diatom images.Then,to address the problem that the multi-head self-attention mechanism can only compute feature maps of a single size,it is proposed to introduce a non-size normalization strategy at the multi-head self-attention module,and construct a feature extraction backbone network NSN-Bo TNet-101 that can compute feature maps of different sizes,to extract richer image features.(3)A few-shot diatom detection model is proposed that combines the NSNMHSA(Non-Size Normalization Multi-Head Self-Attention)and the OHEM(Online Hard Example Mining).In order to solve the problem of low detection accuracy caused by complex diatom image background and diatom object occlusion,a few-shot diatom detection model NMSOFDD(NSNMHSA and OHEM Few-shot Diatom Detection)is proposed,which is based on the two-stage fine-tuning method TFA,using NSN-Bo TNet-101 as the feature extraction backbone network,and introducing OHEM into the model predictor.And according to the specific situation of few-shot diatom detection,the model training scheme is improved.(4)Through the ablation experiment and comparative experiment analysis of the few-shot diatom detection model,the experimental results show that using the mean accuracy precision(m AP)as the evaluation metric,compared with TFA,the addition of NSNMHSA and OHEM modules improved 1.47 and 3.12 percentage points,respectively,and the simultaneous fusion of both improved 5.89 percentage points,verifying that the NSNMHSA and OHEM modules have the effect of improving the accuracy of TFA respectively,and the detection effect of combining the two is better.And compared with other few-shot object detection models,NMSOFDD achieves a m AP of 69.60%.Compared with Meta R-CNN and FSIW,which are based on meta-learning,the m AP improves by 8and 8.7 percentage points,respectively;compared with De FRCN,which is based on metric learning,the m AP improves by 4.96 percentage points;compared with MPSR and FSCE,which are based on migration learning,the m AP improves by 8.69 and 8.34 percentage points,respectively,verifying that the solution proposed in this thesis performs better in comprehensive detection performance.