Sparse Representation to Localize Objects in Underwater Acoustic Images

Screening of underwater images for the presence of submerged objects and phenomena like submarine volcanoes is based on the spotting and localization of objects present in underwater. This paper proposes a dictionary based approach in which a dictionary is created from various sub-images of the target image. The dictionary of sub-images is created for multiple scales. The coefficients are computed by representing the sub-images as a combination of dictionary elements in a linear fashion for the chosen input image. Using these coefficients the confidence measure is determined. This measure designates the likeliness of the presence of the desired object. The confidence measures are computed independently for red and gray scale images which are coalesced to form a confidence map. Blob detection is a set of methods whose objective is to identify parts in an image which have properties namely colour and brightness different from that of the surrounding regions. The maximum response is expected to be at the location of the desired object. This can be determined by running a blob detector on the confidence map.