Application Of Improved Object Detection Algorithm In Video Disc Detection

Object detection,as an indispensable part of computer vision,has been closely studied.In the traditional method,by establishing a complex model,positioning and classification according to the texture and lighting characteristics of the image,there are problems such as poor generalization and large calculation,which is not suitable for the current large-scale image to deal with.The rise of convolutional neural networks has provided new research methods for object detection.The shortcomings of traditional methods have been made up by using convolutional neural networks.At present,the typical single-stage target detection algorithms based on convolutional neural networks include YOLO(You Only Look Once)algorithm and SSD(Single Shot Multibox Detector)algorithm.SSD algorithm has achieved good results in object detection,but there are still problems such as insufficient understanding of contextual information and loss of deep information,and further research is needed.The optic disc is the area that the doctor must observe in detail when diagnosing and treating the fundus disease.Studying the automatic detection method of the optic disc in the fundus image can assist the doctor to complete a lot of works.In the traditional methods of detecting optic discs,they rely on the characteristics of the optic disc,such as shape,brightness,and blood vessel direction,to construct matching features to detect the optic disc.These methods have a large impact on human factors,feature extraction time is longer,and the efficiency of optic disc positioning Low,so it is not suitable for high-volume,efficient video disc inspection.In view of the above problems,this paper focuses on the two key points of the current SSD algorithm and the application of typical object detection algorithms on the video disc.The main contents of this paper are as follows:Firstly,a single object detection algorithm CP-SSD(Context Perception SSD)is proposed.The CP-SSD algorithm promotes the network's understanding of global information through the use of contextual information and scene awareness modules,thereby capturing feature information of objects of different sizes.Deeply uses the semantic activation module to adjust the interdependence between contextual feature information and channels through self-learning,and enhance useful semantic information.The CP-SSD algorithm was validated on the standard data set PASCAL VOC 2007.The experimental results show that the m-AP of the CP-SSD detection algorithm reaches 77.8%,which is 0.6% higher than the SSD algorithm,and the detection effect is significantly improved in images that are difficult to distinguish between objects and background.Secondly,based on the current advantages of deep learning in object detection,an object detection algorithm based on convolutional neural networks is used to detect the optic disc in the fundus image to improve the accuracy.In this paper,the YOLO algorithm and the improved CP-SSD algorithm are studied on the video disc data set.The YOLO algorithm divides the fundus image input network structure into N × N grids,and each grid detects whether the center of the optic disc falls into the grid.During the prediction process,small-scale feature maps are combined with large-scale feature maps to output bounding boxes of different sizes.All the output bounding boxes are screened by nonmaximum suppression thresholds,and finally the disc area is detected.In the CP-SSD algorithm,the fundus image is first down-sampled,then a context awareness module is added behind the VGG16 network structure to enhance the understanding of the global information of the fundus image.At the same time,a semantic enhancement module is added before the convolutional layer prediction of different scales,so that the network detects the disc more accurately.In this paper,the YOLO algorithm and the CP-SSD algorithm are respectively tested on three publicly available retinal image datasets of DRIVE,DRISHTI-GS1,and MESSIDOR.The accuracy of the optical disc detection of all algorithms is 100%.At the same time,the average Euclidean distance between the center point of the optic disc and the standard center point was detected by experiments at 15.43 px and 22.45 px,and the average time required to detect each fundus image was 0.1s and 0.1476 s.The algorithm based on convolutional neural network was verified efficiency and accuracy of the method for detecting optic discs.