Defense and Maritime Solutions

Transcription

1 Defense and Maritime Solutions Automatic Contact Detection in Side-Scan Sonar Data Rebecca T. Quintal Data Processing Center Manager John Shannon Byrne Software Manager Deborah M. Smith Lead Hydrographer 22 February 2012 NATIONAL SECURITY ENERGY & ENVIRONMENT HEALTH CYBERSECURITY

2 Side-scan Sonar Data Side-scan data are still collected regularly, particularly in shallow water Processing Manual Time-consuming Prone to operator error and missed object detection 2

3 Side-scan Sonar Data Side-scan data are still collected regularly, particularly in shallow water Processing Automated Save labor hours As good as a trained hydrographer 3

4 Side-scan Sonar Data Side-scan data are still collected regularly, particularly in shallow water Processing Automated Save labor hours As good as a trained hydrographer Detect 95 percent of objects 1m 3 or larger No more than 5 percent false detection rate 4

5 Automatic Contact Detection (ACD) Proof of concept Pd = 93% and Pfa = 6% Automated Side-scan Data Analysis. Hydro International, Vol. 11, No. 9, p.22-25, October Automated Side-Scan Target Detection: New Tools for Hydrography, Proceedings of the Canadian Hydrographic Conference and National Surveyors Conference, May Alpha results Pd = 90-95% and Pfa = 8% Results and Metrics from an Automated Approach to Detection of Contacts in Side-Scan Imagery, U.S. HYDRO 2011 Conference Proceedings. Beta results This presentation 5

6 Process Flow - Detection Detector finds a peak followed by a shadow Uses split window normalization - constant false alarm rate (CFAR) detection Peak score + shadow score = total score If total score is greater than threshold detection is triggered Detector also includes Bottom tracking Two-dimensional sand-wave suppression filter 6

7 Process Flow - Image Processing Images of the side-scan sonar data are produced All further processing is performed on these images, rather than the side-scan sonar records Up to 264 parameters are extracted from both raw and processed images The parameters are normalized and prioritized using the Mahalanobis distance and the covariance Parameters are used to train the neural network classifier Parameters are used for measuring the length, width and height of the contact 7

8 Process Flow - Neural Network A neural network classification scheme is used to manage the high number of false detections created by the detector Determine a classifier that is a model of the data (does not memorize the data) Network activation values are used to classify each detection as either Contact Clutter (false alarm) The network activation for each detection is plotted (circle) along with beta distributions fit to the network activations from the entire training set 8

9 Process Flow - Hydrographer Review Interactive review of the automatically detected contacts Performed with the SAIC SABER Imagery Review program The detections displayed are overlain on the side-scan sonar record and can be color coded based on flags, such as automatic accept/reject, manual override accept/reject, etc. The hydrographer may override detections or measurements as necessary or manually create new contacts 9 SABER = Survey Analysis and area Based EditoR (SAIC)

10 ACD Trial Applied to NOAA Sheet Full-scale trial performed using beta version of SABER version 5.0 Traditional, manual, two-pass side-scan data processing pipeline Apply SABER s ACD pipeline, followed by a hydrographer quality control review Compare and evaluate results from two separate processing runs Beta trial dataset was Sheet 3 (H12338) Klein 3000 data acquired at 50-meter range scale hours of side-scan acquisition over 11 days Depth ranges from 5 to 28 meters Bottom type is medium- to fine-grained sand Side-scan data is moderate quality Refraction Low density of objects proud of seafloor except in fish havens ACD = automatic contact detection Klein = L-3 Klein Associates, Inc., SABER = Survey Analysis and area Based EditoR (SAIC) NOAA = National Oceanic and Atmospheric Administration 10

11 Current Results Data processing trial ACD team did not consult multibeam data during timed review Digitize all contacts proud of the bottom with a height of 0.5 meter or greater Except in the fish havens Manual, traditional processing pipeline produced 110 contacts 44 of these were determined to be significant; correlate to 18 significant multibeam features ACD processing pipeline produced 120 contacts 41 of these were determined to be significant; correlate to 18 significant multibeam features One rejected by human reviewer One not automatically detected on same part of wreck One in data with refraction noise; not detected Results presented here focus on the subset of significant contacts 11 ACD = automatic contact detection NOAA = National Oceanic and Atmospheric Administration

12 ACD Results ACD detector produced 318,185 detections ACD neural network classified 7,885 detections as contacts (accepted) Neural network trained using detection examples from Sheet F, Sheet H, and Sheet R Neural network contained 60 percent accepts and 40 percent rejects (clutter) Neural network activation threshold chosen at 0.90, on scale of 0-1 Rejected detections were deleted. ACD optional filter to remove duplicates in dual frequency data output 7,042 User-selectable filter to chose duplicate to retain. Highest total contact score was used One significant object was not detected in either side-scan coverage (100 or 200 percent) It was digitized by the human reviewer Of 41 final contacts from ACD/human processing 31 were automatically detected and accepted by the neural network One was detected but rejected by neural network Nine were not automatically detected ACD = automatic contact detection 12

13 ACD Results (continued) Neural network is doing great P fa ~ Difficult to fully test in fish havens Correctly accepted 97 percent of significant contacts detected Automatic detector/neural network P d = 0.70 for single side-scan coverage. Of those missed by the detector Three were less than 0.5m in height Four were between 0.5 and 1.0 m in height One greater than 1.0 m in height was at nadir One was missed Automatic detector/neural network P d = 0.92 for single pass significant objects 1m 3 or greater. 17 of 18 significant objects were automatically detected and accepted by the neural network in at least one side-scan coverage (94 percent). 100 percent were captured by the auto/human complete process. ACD = automatic contact detection 13

14 Labor savings hours of side scan data were processed as part of test Computer processing time for ACD 61.4 hours for detector (27 percent of acquisition time) 9.9 hours for neural network to run (4 percent of acquisition time) Achieved approximately 40 percent labor savings Human review of ACD required slightly more effort than traditional second review averages, but a computer did the work of the first reviewer. Additional time for quality control was mainly due to data management Some time spent on false alarms Results of trial suggest potential effort reduction from 40 percent to 50 percent ACD = automatic contact detection 14

15 Summary and Conclusions ACD processing pipeline will provide considerable savings In labor effort required to process side-scan sonar data ACD provides a robust, repeatable, automated side-scan processing workflow Allows single-pass human quality control review Even with low P fa, high number of detections means lots of contacts to be rejected Tradeoff to manage between sufficiently high P d and adequately low P fa All ACD algorithms and user interfaces now included in SABER v5.0 SABER v5.0 product with beta version ACD released in December 2011 Planned field use of official release in survey season ACD = automatic contact detection SABER v5.0 = Survey Analysis and area Based EditoR version (SAIC)

16 Thank You Rebecca T. Quintal, Data Processing Center Manager 221 Third Street, Building A Newport, Rhode Island USA Tel: rebecca.t.quintal@saic.com John Shannon Byrne, Software Manager 221 Third Street, Building A Newport, Rhode Island USA Tel: john.shannon.byrne@saic.com Deborah M. Smith, Lead Hydrographer 221 Third Street, Building A Newport, Rhode Island USA Tel: deborah.m.smith@saic.com Visit us at saic.com 16