Research On Thangka Cultural Element Detection Based On Improved FCOS Algorithm

Thangka is a painting art form with Tibetan characteristics.Its main contents involve religion,history,politics,medicine and other fields,and it is an important part of Tibetan culture.The Thangka cultural element detection is a key content in the digital inheritance and protection of Tibetan culture.Its main task is to accurately and efficiently identify and locate cultural elements of predefined categories from a given Thangka image,including persons and weapons and so on.It can not only provide technical basis for the works like instance retrieval of cultural element,but also provide materials for the applications like Thangka cultural creation,which is conducive to the inheritance,protection,innovation and development of Tibetan culture.However,currently,there are not many researches on the Thangka cultural element detection,and most of them only detect single types of cultural elements such as large Buddha statues,so the algorithms have some shortcomings in practicability.This thesis expands and improves the open source Thangka dataset to make it more suitable for the task of Thangka cultural element detection.Meanwhile,according to the characteristics of Thangka images,this thesis improves the classical Anchor Free object detection algorithm named FCOS and designs an accurate and efficient Thangka cultural element detector.The main work of this thesis includes the following two points:(1)Propose an ambiguous sample processing method based on center ness.Similar to the Anchor Based object detection algorithm,FCOS also encounters the problem of ambiguous samples when dividing positive and negative samples.That is,the training samples are divided either positive or negative according to the specific threshold and the position of feature points,and when the model trains the samples whose feature point is near the borderline,it will incur great loss,which will make the model pay too much attention to these ambiguous samples and reduce the performance of the model.To solve this problem,this thesis proposes an ambiguous sample processing method based on center ness.This method improves the FCOS from three aspects of soft label,classification loss function and weight optimization of samples.By making full use of center ness,the impact of ambiguous samples on performance is mitigated and the accuracy of cultural element detection is improved.(2)Propose a spatial location information fusion method.In Thangka images,the position of cultural elements is affected by many aspects such as the style,cultural connotation and religious significance and so on.For example,cultural elements are usually assembled into a whole element through some fixed drawing methods(drawing constraints),and then placed in specific positions in the Thangka image(layout constraints).Therefore,the element location information is an important prior knowledge in the task of Thangka cultural element detection.In order to make full use of the location information of Thangka image,this thesis designs a spatial location information fusion module for FCOS.This module integrates the location coordinate information of feature points with the input feature information,and then generates the weight parameters required by the spatial attention mechanism to adjust the feature information,so as to assist the detection network to better complete the detection of Thangka cultural elements.A series of experiments are carried out in this thesis to verify the effectiveness of the proposed method.The final experimental results on the test set show that the m AP and Acc of the detector obtained by the proposed method are 61.69% and 74.72%,which is better than some common mainstream detectors(e.g.YOLOv4).And the AP of important culture elements such as person,lotus throne and mandala reaches 76.64%,76.90% and 86.61%,respectively,meeting the needs of the Thangka cultural element detection task.