Research On Staff Removal Algorithm Of Hand-written Music Score Image Based On Machine Learning

Optical music recognition mainly researchs the conversion of musical score image to computer-recognizable semantic symbols.Staff removal is one of the important modules.The quality of the staff removal algorithm directly affects the final result of the optical music recognition system.Especially hand-written score image,staff removal is especially difficult,because handwritten symbols are complex and variant,overlapping with various deformations and noises.Therefore,staff removal of handwritten music score images has important theoretical research significance and practical value.To improve the robustness of the staff removal algorithm,the study proposed a staff recognition and removal method based on multi-scale multi-direction local binary pattern and XGBoost.First,the study designed an improved local binary pattern operator based on the characteristics of music score image.Secondly,multi-scale local binary pattern features were extracted and concatenated into a high-dimensional feature vector.Finally,using the trained XGBoost model to recognize the staff,and then removal the staff.The experimental results show that F-measure is 97.19% on all test data,indicating high Precision and Recall.F-measure is respectively 96.43%,98.36% and 96.79% on 3 different test subsets,indicating good robustness.Compared with the existing lightweight staff removal methods,the proposed one has better performance in F-measure.In order to directly use grayscale music images,a handwritten music staff removal algorithm combining U-net and Res Net is proposed.First,build a suitable network structure and change the U-net bridge to a Res Net jump connection.Then,the original image data is data enhanced to obtain more training data.Finally,the training image is cut into pieces and the constructed deep learning network model is trained.The experimental results show that F-measure is 98.98% on the binary score image,and 99.09% on the gray scale image,compared with the existing non-deep learning methods,the method has a great performance improvement,and compared with other deep learning models,it also has better removal results.