Study at TCU

This study aims to extract or transcribe a sequence of chords from an audio music recording. For many applications in music information retrieval, extracting the harmonic structure of an audio track is very desirable. Recently, various methods using deep learning have been proposed. Most of these methods use traditional features such as CQT and DFT, even though the musical features consist of both logarithmic fundamentals and linear overtones. This study employs the rich features including both logarithmic and linear frequency features, then let the deep neural network select the effective feature to estimate chords. This combination of the redundant features and deep learning achieved the state-of-the-art performance.

Semantic segmentation is one of the key computer vision tasks which performs the pixel-wise object classification. In these days, the deep learning with CNNs has become the standard when dealing with Semantic segmentation. CNNs is basically very powerful but the visual cortex has the fixed square shape, so many layers should be required to segment objects that have complex contour. CNNs, on the other hand, have a bottleneck structure to represent the highly abstract features of an object, which inevitably blurs the represented shape of the object. To solve these problems, we proposed the relatively shallow and multi-resolution network using deformable convolution. This method enables to segment not only large objects but also thin and small objects.