Research On Detection Methods Of Text In Natural Scenes Based On Feature Fusion

Natural scene images have complex backgrounds and large differences in definition.We discuss the improved methods of text detection in natural scene images based on the classic MSER algorithm and EAST model in this thesis.Our work improved the performance of text detection in natural scene to a certain extent.(1)For the problem of low recall and precision of traditional MSER algorithms on images with blurred boundaries and greater lighting effects,we present an improved text detection method based on visual saliency and MSER By fusing the MSER region after edge enhancement and geometric feature filtering with the salient target region obtained by visual saliency detection,we obtained text candidate areas as many as possible.The CNN text discriminant classifier from the training images is used to classify the features so that the background region can be filtered out more effectively.(2)For the problem of low performance in detecting larger texts and longer texts due to insufficient receptive field capabilities by the EAST model,we present an improved EAST-PAN text detection algorithm.By the analysis of the architecture of the EAST model,the PAN structure in the pyramid attention network is fused to the bottom layer in the EAST basic network backbone ResNet-50.Thus,the pixel-level attention information provided by the FPA module in the PAN network enlarges the receptive field of the network and enhances the ability of the network to perceive long texts by means of the pyramid structure of feature extraction process in the EAST model.At the same time,the global average pooling operation takes into account the overall information representation,so that the detection performance has been improved to a certain extent.(3)For the merging strategies of regression text-box in the EAST model,the weighted processing regression text-box merging operation is improved to use the median of the partial coordinates of the regression text-box belonging to the same text-line.And then,the coordinates of the final text-box are recalculated according to the determined coordinates.The improved strategies consider more information of bounding text-box,and increase the probability that the texts near the edges of two ends are detected by the final text-box.The improvement of the text-box merging strategy based on the improved model structure improves the detection performance of long text in natural scene images.