Research on text detection method based on improved yolov3
Huibai Wang, Hongqing Shi · 2021
Text detection in natural scenes is one of the hot issues in the field of computer vision. Aiming at the problems of complex text background, object occlusion and illumination changes in natural scenes, this paper proposes a method of text detection in natural scenes based on improved YOLOv3 (UDSP- YOLO), this method uses the CLAHE image enhancement preprocessing method to eliminate the impact of lighting changes in the natural scene on the target recognition effect, and uses random spatial sampling pooling (S3Pool) as the downsampling method of the feature extraction network to preserve the space of the feature map Information solves the problem of background interference in complex environments. The 8 times down-sampling feature map output by the feature extraction network is up-sampled by 2 times, and the 2 times up-sampling feature map is spliced with the feature map output by the second residual block to establish The output is a 4 times downsampling feature fusion target detection layer, and the improved network model is used to compare experiments with the original network on the ICDAR2015 data set. The results show that the improved network model effectively improves the detection accuracy.