Automatic segmentation and object classification with neural network for an airborne ultrasound imaging system

An airborne ultrasound imaging system was developed for reflection tomography. The ultrasound transducers surround the region of interest (ROI) in an arrangement optimized for maximum coverage and homogeneous distributed image quality. In this work, we developed a workflow for automatic segmentation and classification of objects in the reconstructed images. Our workflow can be applied for varying intensities of object edges with a local maxima based segmentation and a multi-parameter image reconstruction. The segmented regions are classified with a neural network, and the object localization was implemented with Generalized Hough Transform using a custom template for each classified object in the data set. A classification accuracy of 95% for six trained test objects and a localization accuracy of 5 mm were achieved.