Research On Machine Vision Method In Multimedia Lcd Screen Detection

This paper implements a method for enhancing the reliability of real-time recognition of multimedia display screens.At the same time,it is based on image key point features and deep learning means to process in parallel to enhance the accuracy and reliability of recognition results.It can be used for multimedia automated detection with a display interface in.Since the host computer has determined that all possible interfaces are known in software design,the research core of this subject starts from the image retrieval method,that is,the screen target recognition task is retrieved in the sample library,and has the same as the target image.The template category with the highest similarity is regarded as the content displayed on the current screen.First,a similarity retrieval method based on hash algorithm and color histogram is proposed.By defining several methods to measure the similarity between images to complete the search and target,the known template class that can form the best matching display interface is used as the prediction category,but the final experiment The data shows that the input image is obviously subject to the problems of insufficient stability and low recognition rate after slight disturbance.Then an image similarity retrieval algorithm based on geometric registration of image stable feature point sets is designed.Feature points are extracted from target images and reference templates and matched,and then several feature points with the best matching degree are selected to complete geometric registration.In the calculation and analysis process,when the registered target image and a template have a sufficiently high similarity in the spatial distribution of feature points,the targets can be considered to be consistent,and the template type is the type of the target image.Convolutional neural networks have been widely used in image processing and target detection.Various well-developed network frameworks and auxiliary methods have also reduced a lot of work for applications in various occasions.Through the specific analysis of this task,the application effect of Yolo network is first tried.This end-to-end detection method based on deep learning can basically meet the accuracy requirements,and the speed has reached the real-time level.Reasonable setting of the network output confidence threshold can obtain the detection results with the highest accuracy in the shortest time.However,if abnormal interference or other accidental accidents occur,and the highest similarity is lower than the threshold,in order to avoid recognition errors,the recognition subsystem embedded in the system based on the geometric registration of feature sets can be activated at this time.Because the multimedia detection platform needs to run for a long time and count the number of operations and UI switching at the same time,this security mechanism can guarantee errors caused by unknown factors in a few cases,and thus have an irreversible impact on the test conclusion of a highly automated level.The setting of the double detection mechanism makes the result after the second recognition highly reliable,which is the final recognition result under this period.The above work is completed when the target image is input to the control system transiently,which can greatly improve the system.Recognition accuracy.The final result shows that this multi-core reliability enhancement method further enhances the flexibility and adaptability of the system.When the relative height,position,angle of view or light intensity of the camera changes greatly,even a small part of the screen temporarily obstructs Under the same conditions,the test accuracy can reach more than 99% under the same conditions,and the speed can be increased to 20 fps,which fully meets the system requirements.