Pixel-wise TV Logo Recognition Algorithm Based On Skip-layer Deep Convolution Feature

Currently with the rise of various video applications on Internet,such as video on demand or video living broadcast,videos through Internet spread are turning up.The videos are easy to be artificially tampered with,thus their qualities are uneven.In the face of massive videos,it's inefficient to classify them through traditional manual methods.So it's urgent to provide efficient technical support for related platforms.As an important sign of a video,TV logo reflects video source,category,program orientation,thus by identifying TV logo can achieve accurate video classification.The aim of this paper is to solve TV logo recognition in video.It's a typical fine object recognition problem:TV logo region is small and contains few information.Besides,Hollow-out and translucent logos are easily influenced by background in video frame.Referring to above problems,a pixel-wise TV logo recognition network PNET based on an end to end fully convolutional deep neural network is proposed.First we construct a fine-grained TV logo dataset containing an image set and a label set.we get an image set by extracting and preprocessing video frames;And we get a binary label set by proposing an image by image pixel-wise semi-automatic annotation method.Then we propose a PNET based on a typical classification network AlexNet or VGGNet.We combine global information from deep layers and local information from shallow layers by introducing a skip-layer architecture.We convert model parameters learned by a classification network in a classification task to model parameters required by PNET in a segmentation task by fine-tuning.Finally,taking arbitrary size test image as an input,well-trained PNET will output the same resolution pixel-wise segmentation result.Experiments show that PNET achieves accurate pixel-wise segmentation.The accuracy is up to 98.3% and inference time for per image on 24 G NVIDIA Tesla K80 is less than 1.5 seconds.In this paper,TV logo automatic recognition by deep convolution neural network can provide rich semantic information for video content analysis.Therefore,PNET can be extended to content-based video image retrieval,video classification,commercial advertising monitoring,ratings statistics and other application platforms.