Diffwave, a diffusion based model, generates high-quality stable audio for noise cancellation

Conditional DiffWave uses pressure-affected anti-noise and a novel frequency-weighted loss

Low- and mid-frequency focus improves ANC comfort by reducing pressure-related effects

Combining original and novel losses ensures stable training and mitigates ANC pressure

Tests show strong ANC pressure reduction and waveform recovery, though variability remains

Proposal of an indoor location estimation pipeline that uses a single RGB image as input.

Consists of camera anchor detection, monocular depth estimation and distance scaling, and circle intersection-based coordinate calculation.

The anchor detection stage showed stable generalization performance, and the distance scaling stage effectively restored distance patterns like ground truth through Random Forest regression.

In the final location estimation stage, it achieved an error of 100–130 cm on average, ensuring practical accuracy.

Driver Emotion-Aware Personalized AI Music Composition

Generate Latent vector about melody, rhythm, chords, musical instruments

Build pattern-based classification model and relatedness-based clustering model

AI composition music based reinforcement learning

Generated music can be played longer, perform Reinforcement learning 

Using Double DQN or Dueling DQN

Build Music Relationship music Regressor for arrangement recommend and Emotion-based search

Define the MRF Score, that indicates the relationship between music

Classification about lightness, saturation, and color mean user's emotion

Define loss function about pitch and density in music

After pre-learning latent vector z in Music VAE, training z using Actor-Critic method with loss fuction

Music Source Separation & Automatic Music Transcription using Deep Learning

Fashion Component Classification using Mask R-CNN 

& Fashion Image Generating using GAN

Indoor position recognization under BLE RSSI enviroment using Deep Learning

(CNN, Auto-Encoder, ...)