master topics digital signal processing and image analysis
TRANSCRIPT
Convolutional networks for imbalanced classes
• Focus: study how to improve performance of deep learning with imbalanced classes– Performance measures– Stratified sampling– Using alternate cost functions– Treat as anomaly detection– Other options
• Supervisor: Anne Solberg
Muscle size & architecture during locomotionWith Oliver Seynnes, NIH
Example of muscle contractionThis is a scan of the vastus lateralis and vastus intermedius muscles.Typically, fascicles shorten and their curvature is altered.
VL
VI
Remote sensing: forest classification based onlidar and hyperspectral images
Unsupervised pixel classificationFalse colours display from hyperspectral image
Natural colour channels from hyperspectral image(a) Vegetation height in lidar image
In cooperation with Norsk RegnesentralMain supervisor: Anne Solberg
Typical road, aerial photo
Remote sensing: detection of cars in aerial/satellite images
Detect moving cars forpollution monitoringChallenges:- Parked cars- Shadows
Deep learning possible
In cooperation with Norsk RegnesentralMain supervisor: Anne Solberg
Satellite image, Tanzania
Remote sensing: estimating tree height from shadows Using Sentinnel-2 data (10m resolution)Remote sensing: estimating tree height from shadows Using Sentinnel-2 data (10m resolution)
Challenge: how to get robust estimatesin low resolution and noise
In cooperation with Norsk RegnesentralMain supervisor: Anne Solberg