Research On Traffic Sign Text Detection Algorithm For Natural Scenes

Since convolutional neural network was deployed in computer vision successfully,the field of computer vision has made great progress.Compared with traditional methods of artificially extracting features,convolutional neural network has more powerful feature extraction capabilities and regression fitting capabilities,thus it can meet higher performance requirements.In the field of computer vision,many research work is carried out based on public data sets,and the actual scene is far more complicated than the scenes of public data sets.Therefore,it is often necessary to correct existing algorithms for specific needs.In this paper,a multi-level traffic sign text detection algorithm for real road condition scenarios is proposed,which is improved on the basis of the general object detection algorithm and meets the practical requirements of the industry.The specific content can be divided into two parts.The first part is the design of the traffic sign text detection algorithm,which is designed to design a detector for detecting text in the traffic sign area.Specifically,the traffic sign text detection algorithm is developed on the basis of SSD and RetinaNet.This paper first improves on the basis of SSD,and then conducts comparative experiments on the improved settings to verify the significance of the improvement.Finally,the improved method was migrated to RetinaNet,and better experimental results were obtained.The improvement work of this part is mainly divided into two points.The first point is that the IOU(Intersection-Over-Union)calculation method in the NMS(Non-Maximum Suppression)stage of the general object detection algorithm cannot suppress the phenomenon of the frame-inner-frame under the condition of ensuring high detection precision,therefore changing the IOU calculation method to IOB(Intersection-Over-Bounding Box)calculation.The second point is for the multi-scale feature prediction method of SSD,different aspect ratios are set for different prediction branches according to the different receptive fields and the aspect ratio of ground truth frame in different scales of the data set.Compared to the TextBoxes method,the detection effect is equivalent,but the number of priori boxes is greatly reduced.The second part is the design of multi-level object detection algorithm.This part aims to cascade traffic sign detection algorithm and traffic sign text detection algorithm to form a multi-level traffic sign text detection algorithm flow.Sometimes the area occupied by the traffic sign text area is very small compared to the original picture that was collected under the actual road conditions.If the single-level text detection algorithm is directly applied to the original image,the detection performance will be greatly reduced.Aiming at this problem,the multi-level traffic sign text detection algorithm designed in this paper will firstly detect the traffic sign of the original image,then make projective transformation to the traffic sign detection result,and then carry out the traffic sign text detection according to the results of the projective transformation,and finally the result of the traffic sign text detection is mapped back to the original image to obtain the final detection information of the traffic sign text.