Research On Fine Identification And Retrieval Of Criminal Investigation Images

When a case occurs,a large number of relevant case images will be entered into the physical evidence database as evidence.How to retrieve highly similar images in the database according to a query image has a great application value to handling a case.Content-based image retrieval(CBIR)returns a list of similar images by calculating the distance between the two feature vectors.However,the quality of the extracted image is not high because of the cluttered background and ambiguous subject object,which affects the retrieval effect.In recent years,deep learning technology has made a breakthrough in the field of computer vision.Deep convolution neural network is used to achieve representation learning,then the image features are adaptively extracted and performing retrieval tasks,which has achieved better results than traditional methods.Based on the deep learning technology,this paper explores the methods of Criminal Scene Investigation Image retrieval.The main work of the study includes:(1)Researched the extraction method of image global features.The advanced and mature VGG and Res Net network structures are introduced as the backbone networks for feature extraction.Based on the idea of transfer learning,the model parameters for pre-training on Image Net are imported,and th e feature extraction networks suitable for this datasets are obtained by training and fine-tune on the forensic image datasets.Based on this network,the features of convolution layer and fully connected layer are extracted respectively to test their performance on the image retrieval task.Aiming at the problem that the feature dimension of convolution layer is too high and the key feature is not prominent,pooling technology is introduced to reduce the dimension and channel attention mechanism is introduced to highlight the influence of the key channel.Finally,by using the complementary feature of multi-layer CNN,the convolution layer features and full connection layer features are fused to perform the retrieval.(2)Aiming at the problem of cluttered background and inconspicuous main target,Faster R-CNN network is introduced to detect and recognize the effective target region in the image,and the feature of target region is extracted to perform the retrieval,which overcomes the interference factors.O n this basis,We test three kinds of fine retrieval schemes: "semantic-local feature","global feature-local feature" and "semantic-global-local feature",and prove that the latter is more effective.(3)According to the actual needs of image management and retrieval in criminal investigation,the image management and retrieval system in criminal investigation is designed and developed based on Web technology.Using the research results of this paper,we build a web API for image feature extraction and retrie val.Calling these APIs in front of the application to realize the core functions of the system,apply the research results to engineering practice,and provide help for police handling cases.