1. Introduction
A self-similarity matrix (SSM) is an important mid-level representation for music structure analysis, which is generated by computing frame-to-frame similarity. The repetition of segments typically leads to diagonal stripes in the SSM [1]. The stripes can be emphasised in the path-enhanced SSM by applying diagonal smoothing to the SSM [2] or to the recurrence plot [3]. As shown in Figure 1, the stripe patterns clearly illustrate the repetition of segments (e.g., verse or chorus) and the repetition within a segment (e.g., the bridge in Figure 1). The stripes of the same group can be generated by shifting a specific structure pattern. Based on this observation, we propose to decompose the path-enhanced SSM into structure patterns (as shown in Figure 2(a)) and shift activations corresponding to each pattern using non-negative matrix factor 2-D deconvolution (NMF2D) [4]. We then construct a new, block-enhanced SSM by computing the SSM of the normalised activations. In order to obtain a robust result, we fuse block-enhanced SSMs obtained with different parameters. Then we detect boundaries on the fused, block-enhanced SSM using a checkerboard kernel [5], and label the detected segments by comparing each pair of segments.