Detection method of Tangut based on dynamic convolution and adaptive feature fusion

Due to the complex strokes, high structural similarity, and the problems of missing characters and fading in ancient books written in Xixia, its detection is still relatively difficult to research. In view of the fact that there are few existing studies on the detection of ancient books written in Xixia and most of them have issues including inadequate detection accuracy, failures in character detection, and incorrect detection, a Xixia detection method based on ODA-DBNet based on dynamic convolution and adaptive feature fusion is proposed by comprehensively analyzing the existing mainstream text detection methods. Firstly, the method replaces the original convolution module (Conv) with the full-dimensional dynamic convolution module (ODConv) and combines it with the ResNet-50 feature extraction network to adaptively adjust the feature scale to obtain multi-dimensional complementary features. Meanwhile, ASF was introduced in the feature fusion stage to solve the problem of inconsistent feature scales at different levels during the phase of feature fusion, and the CBAM module was used to replace the original spatial attention module (SAM) to enhance the representation of multi-scale changes, so as to improve the accuracy of the detection of ancient books of the Western Xia Dynasty. Compared with the basic model, the accuracy and recall rate of self-built Tangut text data set are improved by 3.1% and 2.9%, respectively. Compared with other text detection models, the accuracy and recall rate are improved to a certain extent, and it has universality and can accurately detect the Xixia characters.