Automate Detection and Weight Estimation of Fish in Underwater 3D Images

Abstract

Being able to determine the weight of fish is important information for fish breeding facilities. Current methods rely on manual measurements, but it is interesting to look into how underwater imaging can be used to automate these measurements.

Underwater imaging is a complex problem due to light attenuation and poor water quality yielding lots of light scattering from an illuminating light source. In collaboration with Trollhetta and with the use of data from a cutting edge range-gated camera, we will tackle the problem of automating the measurement of the weight and size of atlantic salmon (Salmo salar) swimming freely inside a fish tank. The camera is manufactured by SINTEF Digital in collaboration with Odos Imaging and Bright Solutions, and illuminates the scene by the use of green light with a wavelength of 532nm.

An Active Shape Model implementation is used as a backend to handle the problem of segmenting salmon in the given images. The segmentation algorithm detects the contour of the fish, and makes it easy to find both the length and height of the fish in image space.

The segmented image is matched against the corresponding depth component of the image and the mean of all pixels is used as a measurement for the distance from the camera to the fish. By using the principal of similar triangles in the pinhole camera model we project information like length and height from image space into world space.

The method proposed in this thesis can be regarded as a proof of concept, and a baseline for further research into this particular problem.