Research on intelligent retrieval technology of digital music database based on deep learning

Traditional music information retrieval technology relies on keyword matching and metadata filtering, which is difficult to meet the complex and diverse retrieval needs of users. In this paper, an intelligent retrieval technology of digital music database based on deep learning is proposed. By constructing an efficient deep neural network (DNN) model, automatic extraction and intelligent matching of music features are realized. Research methods include multi-dimensional feature extraction of audio signal, DNN architecture design and multi-task learning optimization strategy. The experimental results show that the accuracy of the proposed model in music classification task is 91.8%, and the F1-score is 90.3%, which is significantly better than the traditional method and the deep learning baseline model. In the cross-data set retrieval task, the mAP@10 index of the multi-task learning version model is improved by 5-15% among different data sets, showing stronger generalization ability. In addition, the analysis of ablation experiments shows that the introduction of BiLSTM, attention mechanism and multi-task learning all contribute significantly to the performance improvement.