| Text detection is one of the most important tasks in the field of computer vision which has been widely used in automatic driving,image retrieval and other fields.However,text detection in natural scene is easily disturbed by external environment and the orientation and size of text in natural scene are arbitrary,which will affect the performance of text detection algorithm.To solve the problems and challenges faced by text detection in natural scenes,the main research contents and contributions of this paper are as follows:(1)In view of the problem of text that orientation is arbitrary and the problem of small-sized text is easy to miss detection or falsely detection,an anchor-based text detection method has been proposed.This method improves the effect of small-size text detection by using a feature fusion network with fine sampling(SF-ResNet50)to obtain more feature information and anchors.The method also uses the affine transformation to map region proposals with different orientations and sizes generated by the region proposal network onto the feature map,and then uses the bilinear interpolation algorithm to eliminate the error caused by quantization for pooling,which solves the rotation Region-of-Interest(RoI)pooling problem and improves the accuracy of text detection.(2)In view of the problem of text size differences,a pixel-based text detection method is proposed to achieve more accurate adaptive multi-scale text detection.Since the pixel-level features are fine-grained features,which help to improve the precision of detection,this method abandons the way of manually setting anchors and predicts the text bounding box corresponding to the pixel which is determined to be positive sample by directly regressing the pixel.The number of region proposals is reduced and the efficiency is improved.In order to alleviate the negative impact of text size changes on detection,this method divides text instances according to size and assigns them to different feature maps for prediction,so that the entire model can have a good detection effect for large,medium and small texts.In order to perform text detection and text recognition simultaneously,the method also establishes a recognition branch in parallel with the detection branch to recognize the content of the text sequence.(3)This paper combines the image haze removal algorithm using dark channel prior based on guided filtering,target detection algorithm,text detection and recognition algorithm,designs and implements a multi-functional detection system in natural scenes.The system has the ability to resist the interference of fog or haze weather.Inputting an image to the system can not only detect the position and category of various objects in the image,but also detect and recognition the text in the image. |
