| At present,the video processing unit in TV chip includes color adjustment,sharpening,noise reduction,detail enhancement and other modules.VPU plays an important role in improving the quality,details and color of TV picture.However,the adjustment of TV picture quality usually requires a set of VPU control parameters to adapt to all kinds of pictures,which is difficult to achieve the optimal quality of the pictures in all scenes.In order to achieve the common optimization of all scenes by the corresponding image quality processing for different scenes.In the paper,a method of combining artificial intelligence,image classification and VPU control parameters is proposed to realize the common optimization of video scenes.In this method,the convolutional neural network in deep learning algorithm is used for scene recognition of video images and pictures are matched the optimal PQ parameters according to scene categories to achieve the optimal picture quality.The main contents are as follows: partition for video scene,production of data set according to the video scene,model building of convolution neural network for scene classification,model training,testing,optimization and model compression,hardware testing on the hardware platform with usability testing and reasoning accuracy testing,real scene response,PQ parameters matching,and achieving the enhancement of the video image quality.The research process and results of this paper are as follows:1.Divide 18 common scenes from video scenes as image categories for scene recognition,and obtain and classify 320,000 images for scene classification data set by means of video extraction,web crawler and open source data set.The focus was on the recognition effect and accuracy of 6 classifications,such as Skin,Blue,Green,Colorful,Architecture,and Nightscape.And PQ image quality enhancement was only done for these 6 scenes.2.Through theoretical analysis and experimental comparison,the lightweight network model Mobile Net V2 was selected as the most appropriate network framework.A classification network is established on this framework.The best training methods of optimal training super parameter values and model fine-tuning were determined through experimental comparison.During the training process,the floating point training and quantization training were mainly performed,and the size of the scene recognition model obtained was 2.3MB,the reasoning speed was 4.0 milliseconds/frame,and the accuracy was 90%.3.In order to further accelerate the model inference speed,the model compression training combining model pruning and network quantization was carried out without affecting the accuracy of the model,which effectively increased the model inference speed to 2.0ms/frame,and reduced the model size to 2.0MB.The model accuracy was stable at90.83%.4.Embed the network model into the TV chip to achieve real-time scene response on hardware platform and automatically macthed the corresponding PQ parameters to improve the saturation,brightness,contrast of the image and enhanced the details of the dark scene,etc.The improvement of its color saturation,clarity and details can be clearly seen in the processed video.In this paper,the problem that the VPU unit cannot achieve common optimization for different scenes was solved.The video scene recognition is carried out through the CNN network in artificial intelligence,and dynamic adjustment of the video image quality is realized according to the recognition results.It has achieved the effect of co-tuning the image quality of six scences,such as: skin,blue,green,colorful,architecture and nightscape.In the future,we can continue to improve the PQ strategy on the basis of this research,such as: adding more image quality parameter processing solutions,using more images to enrich the data sets and adding more classification to realize the processing of more scenes.This research has certain reference significance to other researches on video image processing using artificial intelligence technology. |
PDF Full Text Request