Performers’ copies of musical scores are typically rich in handwritten annotations, which capture historical and institutional performance practices. The development of interactive interfaces to explore digital archives of these scores and the systematic investigation of their meaning and function will be facilitated by the automatic extraction of handwritten score annotations. We present several approaches to the extraction of handwritten annotations of arbitrary content from digitized images of musical scores. First, we show promising results in certain contexts when using simple unsupervised clustering techniques to identify handwritten annotations in conductors’ scores. Next, we compare annotated scores to unannotated copies and use a printed sheet music comparison tool, Aruspix, to recover handwritten annotations as additions to the clean copy. Using both of these techniques in a combined annotation pipeline qualitatively improves the recovery of handwritten annotations. Recent work has shown the effectiveness of reframing classical optical musical recognition tasks as supervised machine learning classification tasks. In the same spirit, we pose the problem of handwritten annotation extraction as a supervised pixel classification task, where the feature space for the learning task is derived from the intensities of neighboring pixels. After an initial investment of time required to develop dependable training data, this approach can reliably extract annotations for entire volumes of score images without further supervision. These techniques are demonstrated using a sample of orchestral scores annotated by professional conductors of the New York Philharmonic Orchestra. Handwritten annotation extraction in musical scores has applications to the systematic investigation of score annotation practices by performers, annotator attribution, and to the interactive presentation of annotated scores, which we briefly discuss.

中文翻译：

---

乐谱图像手写注释的启发式和监督方法

表演者的乐谱副本通常包含丰富的手写注释，这些注释记录了历史和制度上的演奏实践。手写分数注释的自动提取将促进交互式界面的开发，以探索这些分数的数字档案，并对它们的含义和功能进行系统的研究。我们提出了几种从乐谱的数字化图像中提取任意内容的手写注释的方法。首先，当使用简单的无监督聚类技术来识别指挥家分数中的手写注释时，我们在某些情况下显示出令人鼓舞的结果。接下来，我们将带注释的乐谱与无注释的副本进行比较，并使用印刷乐谱比较工具Aruspix，恢复手写注释作为对干净副本的补充。在组合的注释流水线中使用这两种技术可以从质量上改善手写注释的恢复。最近的工作表明，将经典的光学音乐识别任务改组为有监督的机器学习分类任务是有效的。本着同样的精神，我们将手写注解提取的问题提出为监督像素分类任务，其中学习任务的特征空间是从相邻像素的强度得出的。经过最初的时间开发可靠的训练数据所需的时间后，此方法可以可靠地提取全部分数图像的注释，而无需进一步的监督。这些技术通过使用由纽约爱乐乐团专业指挥家注释的管弦乐乐谱样本进行演示。乐谱中的手写注解提取已应用于表演者对乐谱注解实践的系统研究，注释者的归因，以及被批注乐谱的交互式演示，我们将对此进行简要讨论。