A handwritten music score spectral line deleting method combining U-net and ResNet

Abstract

The invention relates to a method for deleting spectral lines of handwritten music scores of U-net and ResNet, and comprises the following steps: the U-net and the ResNet are combined to construct a deep learning network structure, and a model suitable for deleting the spectral line of the handwritten music score is established; wherein the input of the deep learning network is an original music score image, the true value label is a corresponding no-spectral-line music score image, the output of the deep learning network is a result image containing a probability value, a final threshold value is set according to specific data, binarization operation can be completed on the result image, and a final result is obtained; data enhancement; and training the model: training the constructed model by using the images of the training set, and carrying out parameter fine tuning according to the data of the verification set to obtain the optimal model parameters.

Description

technical field [0001] The invention relates to the field of handwritten music score recognition, and completes the task of deleting lines of handwritten music scores through technologies such as image processing and deep learning. Background technique [0002] Music score is a music recording method that expresses the sound characteristics of music, such as pitch, interval, beat, etc., through visual marks. The existence of musical scores allows music to be communicated worldwide, and it is also a "textbook" for music lovers to learn. Before the widespread use of printed scores, a large number of musical works were preserved in the form of handwritten scores. However, handwritten sheet music is easily damaged and there is a risk of loss. With the popularization of computers, the speed of information exchange has been greatly improved. At this time, the spread of handwritten music became very slow. Therefore, it is necessary to convert handwritten scores into digitized i...

AI Technical Summary

Problems solved by technology

However, with the deepening of the network structure, it also adds great difficulties to training and learning.

The deep network makes the training time longer. In the case of limited resources and large data sets, it is difficult to achieve fast training.

At the same time, in the process of gradient backpropagation, problems such as gradient disappearance or gradient explosion are prone to occur, which brings great obstacles to training.

Moreover, as the network deepens, the performance of the network may not continue to improve, and may even become worse

Images