Research On Multi-scene Image Text Detection Method Based On Fully Convolutional Network

This thesis studies how to accurately detect text areas in images.Based on the idea of image segmentation,an improved full convolutional network model is proposed,which can effectively detect text areas in different scene images.Aiming at the difference in character characteristics between the network synthetic image and the natural scene image,two two-stage text detection methods based on the full convolution result are designed to achieve the text detection task in multi-scenario images.The main work of this paper is as follows:1)This paper improves the traditional full convolutional network model,and designs a U-shaped full convolutional network model based on batch normalization.Based on the original FCN network,it is reconstructed by hole convolution.Improved methods such as full convolution normalization optimize the traditional full convolutional network model.2)In order to accurately detect fine-grained text in network composite images,this paper proposes a two-stage image fine-grained text detection method FMN.First,a full convolutional network is used to detect and label the text-like regions in the image,and then for the full volume For the problem that the output of the product network is not fine enough,the MSER + NMS algorithm is used to extract fine-grained features for text-like regions.At the same time,the ellipse fitting method is used to detect features such as text character tilt and morphological differences,and finally achieve image text detection.3)Aiming at the problem of text detection of images in natural scenes,this paper proposes a three-stage text detection method based on the output results of full convolutional networks.Considering that the background of the image in the natural scene is too complicated,based on the full convolution output result,a modification of the MSERs tree structure is proposed to eliminate the character sticking in the output result,and combined with image layering,region fusion,and CNN character discrimination,etc.Text detection target in natural scene.4)In order to better describe the image text detection effect,the concept of image text detection granularity is introduced and a mathematical definition is given.Experiments on benchmark data and real data sets show that compared with existing methods,the method in this paper has significantly improved detection granularity and accuracy.