Enhancement, summarization and analysis of videos of Nephrops habitats
Citation:
Ken Sooknanan, 'Enhancement, summarization and analysis of videos of Nephrops habitats', [thesis], Trinity College (Dublin, Ireland). Department of Electronic & Electrical Engineering, 2014, pp 190Abstract:
Harvesting the commercially significant lobster, Nephrops Norvegicus, is a multi-million dollar industry in Europe. Stock assessment is essential for maintaining this activity but it is conducted by manually inspecting hours of underwater surveillance videos. The motivation for this thesis is to improve this manual inspection process by exploring object recognition techniques for detecting these burrows automatically. As the visibility in these videos is generally poor, the recognition aspect of this system is combined with two additional preprocessing steps of image enhancement and content summarization. These three techniques are discussed separately in the thesis.
In the first step of image enhancement, the radial degradations (vignetting) associated with the illumination distribution of the light source and the absorption from water in these images are corrected. To perform this correction a novel image enhancement technique is developed that uses ideas from the vignetting and underwater correction literature. In this technique, a new degradation model is derived to take into account the spatial deteriorations in each colour channel that occur outside of the light beam footprint on the sea floor. Unlike current techniques in the vignetting literature, this model does not restrict the shape of the degradations to being circular and located at the image center, but instead follow a general elliptical shape and center of the light beam footprint. Novel techniques are also developed for estimating the parameters for the model, which use the attenuation from corresponding points across multiple frames. Correction is performed by attenuating pixel values according to the gain field parameterized by the model. When evaluated against a state of the art vignetting technique, the method achieves superior results.
In the content summarization chapter of this thesis, the tedious manual process that scientists endure by inspecting thousands of video frames is reduced to the scanning of a single image. This particular image is a mosaic that is created by aligning and rendering all of the video frames together. These mosaics are a useful analysis tool as they offer a wide area view of the surveyed seabed area, which makes it easy for scientists to spot spatial relationships among bm-rows. To align the video frames, a Bayesian framework for registration is developed that uses the burrows (blobs) in these images for feature matching. Once aligned, the overlapping regions among these frames are rendered with a new technique that uses the estimate of the light beam center to capture image details of well-lit regions in the generated mosaic. Experiments performed in this chapter show these new alignment and rendering techniques achieve improved results when compared to four state of the art systems from the literature.
The final algorithm presented involves identifying burrows automatically from the generated video mosaics. Using mosaics for this application improves on the existing video based technique by summarizing the results in a single image as opposed to inspecting thousands of video frames. Recognition in this system is performed by first detecting candidate objects and then classifying
Author: Sooknanan, Ken
Advisor:
Kokaram, AnilPublisher:
Trinity College (Dublin, Ireland). Department of Electronic & Electrical EngineeringNote:
TARA (Trinity’s Access to Research Archive) has a robust takedown policy. Please contact us if you have any concerns: rssadmin@tcd.ieType of material:
thesisCollections
Availability:
Full text availableMetadata
Show full item recordLicences: