WP3 “Automatic Recognition of Mining Targets” focuses on technologies for automatic identification of mining targets in acoustic data, oriented towards general mapping of the survey area and automatic detection and localization of targets using optical information for close range studies.
For acoustic detection of Mn-nodules a multibeam echo sounder is used. Both, the bathymetry and backscatter provided by this device can contain valuable information. The proposed strategy aims to find the largest area of high Mn-nodule likelihood. Additional working constraints of the AUV are considered, too. Four different features are used for determining the promising Mn-nodule areas.
The visual detection of Mn-nodules is based on a deep learning approach. The first step is to perform an adequate image pre-processing on the acquired images. This is necessary due to inhomogeneous illumination and other underwater effects like backscatter and attenuation. For these tasks we perform lighting inhomogeneity correction, noise reduction with bilateral filtering and image dehazing.
In this project, we compared the performance of Fast-RCNN with DarkNet Yolo. Both algorithms are general purpose object detection schemes with fast and accurate implementations. We obtained good results using both algorithms, due to the high frame rate that Yolo is able to rich, we will use it on the AUV for real time manganese nodule detection