Audience Image Recognition Based On Depth Width Fusion Residual Network

Theater audience statistics is a powerful means to crack down on the cheating of box office,ghost field and lock field.There are three traditional statistical methods: one is to count the number of people from the entrance.There are many uncertain factors in the flow of people,which lead to inaccurate statistics;The second is to count the number of people from the seats,install a human body detector on each seat or collect human body temperature infrared imaging,but there will be foreign body occlusion,resulting in the low accuracy of the number of people statistics;Third,the use of cameras to visually count the number of people.The sitting posture and light of the audience in the cinema are constantly changing,which leads to the low stability of the statistics.So it is very necessary to study the high accuracy and stability of the number of people statistical method to solve the problem that the traditional way can not solve.In view of the demand of audience statistics in cinemas,this thesis conducts a research on the number of people based on convolutional neural network,and the main work contents are as follows:(1)Seat image extraction.In the actual projection of the movie,the position of the video camera installed above the cinema is fixed,and the position of each seat in the image is also fixed,but the image size of each seat is inconsistent,and it is not a rectangle,but a convex quadrilateral.Aiming at the problem of different seat sizes in cinema scenes,a seat image calibration tool is designed.Firstly,an unmanned seat is calibrated manually,and the coordinates of four vertices of the convex quadrilateral of each seat are recorded to generate a seat definition file;Secondly,the definition file and the target image are selected to capture automatically;Finally,the captured image is geometrically corrected.(2)Seat image preprocessing.There are many problems in the images collected in the cinema environment,such as lighting,shooting angle and so on.It is difficult to identify whether there is an audience on the seat.In view of the image quality problems caused by the influence of lighting,environment and shooting angle in the cinema environment,the convolution neural network method is used to restore the image,and the data expansion technology is used to improve the generalization ability of the data set.After image preprocessing,the image maintains high image quality before it is sent to the training model.(3)Judgment on the existence of seat audience.In the process of recognition,due to the influence of the lighting and shooting angle of the cinema,in addition to the two kinds of real judgments of someone and nobody,there will also be false judgments such as luggage mistakenly judged as someone or too bright or too dark light mistakenly judged as nobody.Aiming at the problem of poor recognition rate caused by the environmental impact in the process of recognition,based on the classical residual network,this thesis combines the wide residual network,and designs a deep wide fusion residual network algorithm for theater audience image recognition statistics.The preprocessed single seat image is input into the network model,and six groups of different combination experiments are designed to judge the existence of the audience on the seat.Finally,the test data set is used to count the number of people.The experimental results show that the statistical accuracy of this algorithm is higher than other models,which provides a strong basis for improving the authenticity of box office.