Study On Feature Extraction For Crime Scene Investigation Image Retrieval

The popularity of the internet,mobile phones,and digital cameras has made it easier to take photos,upload photos,and share photos.As a result,massive amounts of image data are produced every day.Efficient management of massive image data and the ability to quickly retrieve the required information from massive image data has become one of the current research hotspots.Relying on the Key Lab.for Electronic Information Processing with Applications in Crime Scene Investigation,ministry of public security,China,this paper analyze the image of crime scene investigation(CSI),and combine the characteristics of the CSI image to improve the algorithm of image feature extraction,optimize the image retrieval method,and improve the efficiency of image retrieval.This paper firstly introduces the current situation of the field of crime scene investigation image retrieval,summarizes and analyzes the shortcomings,and then proposes an innovative crime scene investigation image retrieval mechanism.The innovative points include the extraction of low-level digital features,and the optimization of retrieval results.Finally,based on the Convolutional Neural Network(CNN)model,the middle layer feature extraction algorithm of convolutional neural network and the feature extraction algorithm with low-level digital features are proposed.Nearly 20,000 images of CSI were obtained from Shaanxi Provincial Public Security Department,and 10082 images were selected as the experimental data.Firstly,a DCT-DCT wave texture feature extraction algorithm was proposed.The experimental results show that on the CSI images,the DCT-DCT wave has low texture feature dimensions,fast extraction speed,and high retrieval efficiency.In addition,in order to describe the content of CSI images from different perspectives,the GIST descriptor feature is introduced for the first time to extract the features of CSI images.The DCT-DCT texture feature,GIST descriptor feature and HSV color histogram feature are combined to form a fusion feature to retrieve the CSI image.The experimental results show that the final retrieval precision can be greatly improved by using the fusion feature.Secondly,this paper proposes an optimization algorithm based on the retrieval results.The retrieval algorithm is divided into two steps.Firstly,performing a preliminary retrieval on the query image,and then according to the results of the initial retrieval to determine whether to extract the high-level semantics of the query image,so as to retrieve according to high-level semantics.The algorithm makes full use of the retrieval convenience brought by the high-level semantics of the image,and avoids the retrieval error caused by the misjudgment of high-level semantics of the image.The experimental results on the CSI images show that the algorithm has greatly improved the retrieval precision.Finally,since the deep learning algorithm can combine the low-level features to form a more abstract high-level representation feature,the deep learning model is trained on the CSI image database,and the image features extracted from the middle layer of the model are analyzed and processed to form a new image feature that can express the CSI image well.The experimental results show that the image features extracted from the deep learning model are superior to the traditional texture features,color features,etc.in the retrieval accuracy rate.In addition,by combining the traditional features with the features extracted from the deep learning model,the retrieval accuracy can be further improved.In addition,in order to visualize the retrieval results of the CSI images,a retrieval interface of CSI is designed.The retrieval interface can display the retrieval results under six different image features and the semantic categories of query images.The retrieval interface is not only a visual expression of the retrieval algorithm,but also can help the investigators retrieve the desired image.