Research On The Detection Technology Of Modern Tibetan Printed Matter Layout

Modern Tibetan prints are an important part of Tibetan cultural resources.The text lines in the layout of Tibetan modern prints contain a lot of important information.The effective detection of the text line areas in the Tibetan layout is important for the digital protection of Tibetan cultural resources.The layout of the modern Tibetan prints collected in this article is divided into modern Tibetan books and natural layout Tibetan layouts.In view of the fact that traditional detection techniques cannot effectively solve the problems of low resolution and variable font styles in modern Tibetan layouts.How to effectively detect the layout of Tibetan modern printed matter,this thesis uses two deep neural network target detection methods to respectively detect the text line area in the layout of Tibetan modern printed matter.main tasks as follows:In the first part,in order to solve the detection of text lines at various levels and approximate levels in the layout of modern Tibetan prints,the layout target detection algorithm of modern Tibetan prints based on Faster R-CNN is adopted.First,the detection of Faster R-CNN is introduced in detail.The network structure and algorithm principle,the main idea is that when the region suggestion network extracts the target area of interest,then enters the Fast R-CNN network to obtain the overall feature map of the Tibetan layout,and finally Faster R-CNN determines which category the area of interest belongs to,Train the data set in Res Net-50 network structure to realize the classification and positioning of the text area.In this thesis,by setting the anchor to four scales and five ratios,each pixel on the feature map can generate 20 anchor boxes with different scales and sizes.It is recommended that the network generate a total of 1,000 target areas of interest.Experiments show that the improved Faster R-CNN detection algorithm has a better detection effect on the Tibetan modern print data set than the original Faster R-CNN.Experiments were conducted on different scales of Tibetan modern print data sets,and they all have certain detection effects.Due to the complexity of the Tibetan layout in the natural scene,the detection efficiency on the Tibetan layout data set in the natural scene is lower than that on the modern Tibetan book data set,but the overall detection effect is ideal.In the second part,this thesis also adopts the SSD Tibetan modern print layout detection algorithm to realize the detection of text lines with different size ratios.First,the fully connected layer in the Res Net-50 network is replaced with a convolutional layer.On this basis,an auxiliary convolutional layer is added.By using the feature maps in the 3rd to 7th layer of convolutional layers,a total of five scales are extracted.Feature map.Secondly,each pixel on the extracted feature map is set to generate five scales and five prior frames with different proportions and sizes.Finally,each pixel can generate 25 prior frames with different sizes and proportions.The frames are matched to obtain the category confidence and coordinate position corresponding to each a priori frame,and finally,whether the corresponding prediction frame is a text line target is determined,and its position is returned to make it most likely to be close to the real frame.It is found through experiments that the algorithm achieves a good detection effect on both the modern Tibetan book layout and the natural scene Tibetan layout.