剪枝优化的文本区域检测算法

Text Region Detection Algorithm Based on Pruning Optimization

针对实际场景文本检测中存在由于输入图片多方向导致检测结果误差大、多数文本检测算法只是解决单一的文本区域定位的问题,设计并实现了一种基于剪枝优化的多方向文本区域检测算法,对输入的任意方向的图片进行准确有效的文本检测。首先使用混合剪枝的方法改进VGG16网络模型,实现图片方向快速分类任务,并根据分类的结果对图片进行旋转调整为正向;然后通过级联浅层信息的方法改进YOLO网络结构,实现小角度倾斜文本区域特征提取和预测任务;最后对文本方向预测和文本区域预测两个步骤进行组合,实现输入任意方向的图片均能检测出文本区域位置的效果。实验表明,改进的方法具有较好的检测性能,在ICDAR2013和TDS数据集上,对场景文本检测具有较高的准确率和召回率。

In view of the fact that there are many errors in the detection results due to the multi directions of the input image in the actual scene text detection,most of the text detection algorithms only solve the problem of single text area location,a text detection method based on the coupling P-VGG16 and C-YOLO is designed and implemented,which can accurately and effectively detect the input images in different directions. Firstly,the P-VGG16 network model is improved by using the hybrid pruning method,and the P-VGG16 is used to realize the task of image direction classification. Secondly,the C-YOLO is used to realize the task of text region feature extraction and prediction by improving the YOLO network structure by cascading the shallow information. Finally,the two steps of text direction prediction and text region prediction are coupled,the effect of detecting the position of the text area can be achieved by inputting pictures in any direction. Experimental results show that the improved method has better detection performance. On ICDAR2013 and TDS data sets,it has higher accuracy and recall rate for scene text detection.