| Text detection is a very popular element in current image processing research and is used to locate the position of text contained in an image.Scene text contains important semantic content,and analyzing the text content helps to further recognize the image and understand the scene.However,the background of some images in the text dataset is complex,variable and blurred.For such problems,this thesis proposes an improved algorithm to calculate the atmospheric light value and transmittance to deblur the image.In addition,since text detection is a small target detection,the text itself has less feature information.In view of this,this thesis proposes a scene text detection algorithm based on Faster RCNN for effective detection of small text.The main results achieved in this thesis are as follows:Firstly,in this thesis,we make our own text dataset and propose CALVT algorithm for blurred images,which is used to deblur the training images and reduce the noise in the images in order to obtain clear images.Secondly,to address the problem of small text prone to error and omission detection,this thesis proposes a scene text detection algorithm based on Faster RCNN,i.e.,STDA-Faster RCNN.for feature extraction network Res Net,by introducing optimized depth-separable convolution and replacing point-by-point convolution in depth-separable convolution with group convolution,the number of parameters is reduced and the network runtime is shortened.The number of parameters is reduced and the running time is shortened.Then,the channel blending wash is added after the grouped convolution,which can effectively integrate the text feature information between each group,thus improving the feature extraction ability of the network for small text.In addition,when selecting text candidate frames,the NMS algorithm is optimized in this thesis to improve the recall rate of text detection by resetting the confidence level,which effectively avoids the wrong and missed detection of small text.Finally,this thesis does comparison experiments on the homemade text dataset using the original algorithm and the improved algorithm.It mainly includes the comparison experiment of CALVT algorithm for deblurring processing,the comparison experiment of feature extraction,the comparison experiment of candidate frame design,and the comparison experiment of STDA-Faster RCNN algorithm for text detection.The experimental results demonstrate that the accuracy rate of STDA-Faster RCNN algorithm is 85.3%,the recall rate is 76.4%,and the F-value is 80.6%.Compared with the original algorithm,the accuracy rate,recall rate,and F-value are improved by 5.7%,2.8%,and 4.1%,respectively,with the reduction of the number of model parameters. |
