Scene text detection with improved receptive field and adaptive feature fusion

Regression-based text detection methods are currently the research focuses due to their simple network structure and fast inference speed. However, most of them suffer from limited receptive field of convolutional neural network and simplistic feature-fusing in feature pyramid. As a consequence, the previous algorithms still have many shortcomings, such as difficulty in accurately detecting long texts and the inconsistency across different feature scales. To address these two problems, we first incorporate a densely connected atrous convolutional module into the feature extraction network, accordingly the receptive field is enlarged, and in turn the extraction of high-level semantic information is strengthened. Secondly, we weight and re-fuse the features from different levels of the feature pyramid, which can filter conflicting information at various levels to maintain the scale invariance of features. Extensive experiments have been made on ICDAR2015 and MSRA-TD500 datasets, and the experimental results have proved the effectiveness of the method.