Research And Application Of Natural Scene Text Detection Algorithm Based On Deep Learning

With the increasing maturity of deep learning technology,natural scene text detection based on deep learning has become a research hotspot in recent years,and has been widely used in fields such as unmanned driving systems,multilingual image translation systems,and visually impaired navigation systems.However,due to the complexity of the natural scene image background and the diversity of the text,the natural scene text detection based on deep learning still has the following problems:(1)As the number of deep learning model networks deepens,small-scale text semantic features and text boundaries The problem of dwindling information.(2)The problem of inconsistency in the fusion of features of different scales.(3)The problem of inaccurate division of the segmentation map caused by pixel overlap.In response to the above problems,a series of researches were carried out on natural scene text detection algorithms based on deep learning,and a Scene Text Detection Algorithm based on Attention Mechanism and Feature Fusion(AMFF)was proposed.The main research contents of this algorithm are:(1)Introducing the convolutional attention mechanism in the FPN+PAN structure,shortening the information propagation path while extracting more effective feature information,so that more semantic features and boundary information are retained.(2)The adaptive spatial feature fusion method is adopted for features of different scales to reduce the inconsistency problem of feature fusion of different scales.(3)Use an improved progressive scaling algorithm to alleviate the problem of inaccurate segmentation of the segmentation map caused by pixel overlap.In order to further improve the performance of the AMFF algorithm,an improved algorithm based on AMFF is proposed: AMFF Algorithm based on Text Context Feature(TCF-AMFF).The improvements are:(1)Adopting a non-parameter attention mechanism Replace the convolutional attention mechanism in the original network,without increasing the original network parameters,infer the three-dimensional attention weight for the feature map,and reconstruct and enhance the feature information to reduce text missed detection.(2)Increase the extracted text context feature information,obtain more discriminative features,and reduce text misdetection.(3)In different levels of contextual features,the introduction of a non-participant attention mechanism,combined with multi-scale feature fusion,improves the focusing ability of feature maps,and reduces the impact of background and noise on detection.In order to verify the effectiveness of the algorithm,the TCF-AMFF algorithm is applied to RMB text detection and road sign text detection respectively.To a large extent,it solves the image quality problems of RMB text detection affected by stretching,deformation and corner folding,and improves the accuracy of RMB text detection.The problem that the text detection of the road sign is affected by the text diversity is alleviated,and the accuracy of the text detection of the road sign is improved.