Research On Hysteroscopy Video Abstraction Method Based On Image Quality And Attention

In hospitals or gynecological clinics,hysteroscopy video databases are usually built for case comparison,medical history tracking,medical research and teaching,etc.However,the data redundancy in hysteroscopy video is high,and there are many meaningless frames.In order to improve doctor's browsing efficiency,some scholars use video summarization method to extract representative frames from the original frame sequence in order to simplify the data.However,the accuracy rate of the related methods is low.Therefore,this paper adds image quality assessment on the basis of attention assessment,and improves the accuracy by removing those frames with low image quality score.The main work of this paper is as follows:(1)Pre-processing and extracting the main content of the mirror.In preprocessing,frame sampling of hysteroscopy video and color space transformation are performed successively.Then we design a method to extract the main content of the mirror: for the input grayscale image,binarization,hole filling and projection are carried out successively.Then we precisely locate the effective area of the mirror,and extract the corresponding main content.Experiments show that this method can effectively remove redundant data and improve the computational efficiency of video summarization method.(2)Visual attention assessment.To solve the problem that the number of SURF(Speeded Up Robust Features)feature points,extracted from the original hysteroscopy image,is too small,the corresponding LBP(Local Binary Patterns)map is obtained first,and then the SURF feature points are extracted from it.Experiments show that a large number of SURF feature points can be extracted from LBP images of hysteroscopy images,the main reason is that the SURF feature points are tend to concentrate on the areas with rich texture information,and LBP is a kind of texture feature with strong differentiation.Therefore,the LBP image can highlight the texture information in the original hysteroscopy image.On this basis,we match the feature points of adjacent frames and calculate the attention value.The experimental results show that,in the evaluation of attention,our algorithm is better than the traditional optical flow methods.(3)Image quality assessment and a two-stage video summarization method.In view of the lack of consideration of image quality factors in existing research,this paper proposes a method,with two-stage structure,of hysteroscopy video summarization: firstly,key video segments are extracted through image quality assessment,and then the frames with the highest attention value are extracted from each segment to form the video summary.When evaluating sharpness,the intensity map of sharpness is obtained based on DCT(Discrete Cosine Transform)coefficients,and then binarization,morphological closure operation and hole filling are performed successively.Then the ratio of clear area to the total area of image is calculated as the final sharpness value.Experiments show that this method is better than the method that evaluates image sharpness based on the second distortion.In addition,compared with the traditional method of segmenting attention curve into segments with the same length,extracting key video segments based on image quality assessment can effectively filter out the meaningless frames and improve the accuracy rate of video summarization algorithm.In the experiment,a local database was used,which contained 30 hysteroscopy videos,the according frame rate is 25 frames per second.The corresponding results show that the recall rate,accuracy rate and F-measure of the algorithm(extracting hysteroscopy video abstraction based on image quality and attention)in this paper are 83.33%,51.92% and 63.98%,respectively.Compared with other algorithms,the method in this paper has improved obviously in the aspect of accuracy rate,and the corresponding recall rate and F-measure are also higher.