The current dataset comprises digital images (
pixels) of fish filmed in an illuminated enclosure [
1
]. The walls present a bright background with light intensity of
comparable magnitude to the foreground objects and therefore contrast in
light intensity cannot easily be used to differentiate between
foreground and background. However, the outlines of fish exhibit sharply
changing light levels: the outlines are generally darker than the
regions inside and outside. This can be explained by the fact that the
light source is close to the camera: walls and the flatter central
regions of fish reflect light back, whilst the curvature of the fish
surface near the edges causes light to be reflected away and the
observed intensity is therefore lower.
The intensity gradient can be used to highlight the outlines of objects. These are not the only source of edges: biological features (such as eyes), surface texturing and light reflection create edges inside the outlines whilst shadows and enclosure corners create edges outside the fish outlines. The background noise otherwise gives rise to few edges.
Unfortunately, the use of edges creates some difficulties in this application. The edge features described above are suppressed in some instances and enhanced in others. Factors contributing to this variability include: the inclination of the fish in certain directions; colour variation across the surface of an individual and between individuals; and curvature caused by swimming motions of fish. Biological features such as eyes and fins are also highly variable in terms of visibility and orientation. This high degree of variability necessitates the use of robust methods for dealing with edges.
Thus, identifying fish outlines requires the detection of edges. To
create a gradient image, the intensity image was convolved with an edge
filter given by the gradient of a Gaussian density. We obtained a good
balance of smoothness and detail when using a Gaussian of width
pixel. Processing an image in this manner transforms it from a 2D scalar
field to a 2D vector field. The
absolute edge strength
at a point is defined as the magnitude of the gradient, whilst the
directional edge strength
is defined as the magnitude of the component of the gradient in a
specified direction.
Thresholding the intensity image serves no purpose, because the foreground and background light levels cannot be distinguished even for a single fish. However, in the absolute edge image the background noise is largely confined to the lowest 10% of grey-levels and so we threshold the images at this level which removes almost all the noise, leaving those edges which are associated with fish and wall features.
K De Souza