Audio Event Detrction Based On Structured Information Extraction

With the development of deep learning in the field of audio event detection,audio event detection helps people achieve the growing need for a better life in security monitoring,audio scene analysis,noise pollution detection,and multimedia information retrieval.The information of the audio event is structured,which means that the features of the audio event are structured and organized.The structured information of audio events can be divided into audio structured information of audio events and structured information of audio event tags.This paper mainly analyzes and utilizes structured information through the following three aspects.1.An audio information extraction method based on multi-granularity information fusion is studied.Aiming at the diversity of audio structured information,the visual network is used to analyze the mutual influence of different audio types and different granularity networks.The audio information extraction method has improved the extraction ability of audio structured information.Using 13.92%of the model parameters on the Urbansound8K dataset achieves an AUPRC improvement of 0.8%.2.A method for introducing multi-label relationships based on global constraint loss function is proposedAiming at the problem of lack of effective relationship modeling between labels in multi-label tasks,this paper proposes a multi-label relationship introduction method based on global constraint loss function from the perspective of loss function.When a single label is discriminated separately,the overall information in the audio is used,and global constraint information is introduced to constrain the results of a single label.On the DCASE2020task5 dataset,the Micro_AUPRC of the Coarse task and the Fine task are increased by 0.6%and 0.9%,respectively.3.A method of tag semantic sharing based on multi-task learning is studiedConsidering that the Coarse task and Fine task tags in this data set have a superior-level relationship,this paper studies a tag semantic sharing method based on multi-task learning on the basis of common multi-task,and realizes the information sharing of tag semantic information among multiple tasks.On the DCASE2020task5 dataset,the Micro_AUPRC of the Coarse task and Fine task were improved by 0.8%and 0.9%,respectively.Based on some methods in this study,the results of Coarse_Micro_AUPRC 87%and Fine_Micro_AUPRC 76.17%were achieved on the DCASE2020 Task5 dataset,and the team ranked 3rd in the world and 2nd in the country in the DCASE2020 task5 competition.