|
|
|
Executive Summary Supporting two slides
Recent accomplishments
We have constructed a multi-level classification system for the detection of shallow water mines that includes:
- State of the art image preprocessing and enhancement using:
- Innovations in computer vision
- Wavelet denoising
- Advanced detection techniques that is more robust to background changes
- Expert fusion at various levels which optimizes computational efficiency vs. local image complexity. It incrementally applies more refined background analysis (computationally intensive).
- Decision support user's interface to enhance the performance of a human experts by overlaying machine expert prediction on the original image.
Recent results
6.2, 6.3 Transitions
Overview
In our image enhancement and normalization methods we have succeeded in removing a large portion of background structure resulting from ground vegetation and thus made the task of mine detection much simpler and less computer intensive. This has led to a reduction of an order of magnitude in false alarms and the feasibility of 100% mine detection in difficult images without the need to dynamically adjust the parameters to a new ocean environment.
The results have been achieved by applying an array of experts with different model parameters and image representations. Fusion of these experts is done in such a way that experts which are very computationally expensive are being selectively used for specific regions where their utility is largest.
A side effect of the results is the construction of a graphical users interface which can greatly enhance the work of a human expert by alerting him to specific locations in the image and effectively fusing machine expert information with the sonar image. This project is being transitioned to the AQS 14 effort as well as to unmanned underwater vehicles.
Future goals
Technical Contact
Upon completion of the Sonar mine hunting project (DARPA funding ended September 2000), we have come to the conclusion that current navy sonar performance can be significantly improved by using that advanced signal processing methods; In particular, we refer to performing detailed time/frequency decomposition (using a specially designed mother wavelet) and couple that with the design of a sonar pinging signals and with information fusion from different frequency bands. In under water sonars there is a strong constraint on the frequency of the pinging signal, higher frequencies are more difficult to generate (with strong volume) and they decay faster in water. Lower frequency signals provide lower spatial resolution and thus can not resolve fine details (external and internal) of objects, leading to poor classification and detection of underwater mines. We intend to introduce a powerful approximation to a continuous wavelet transform which can run in much shorter time and lead to the the same temporal resolution. We further intend to increase the temporal resolution beyond the limit indicated by the uncertainty principle using an efficient combinations of different frequency measurements. In the supporting slides, we demonstrate the use of continuous wavelet transform to the reconstruction of the external and internal structure of a Manta mine using dolphin pings. This is done without knowing the specific pinging signal -- a feature that could further enhance the image reconstruction and using only pings every two degrees (a total of 90 pings) while a dolphin sends about 200 pings per second, and thus reconstructs at a much more refined angle resolution. The details of the mine as they appear in the reconstruction are demonstrated via the arrows.