Automated Image Analysis for Fisheries Applications

We propose to develop software that is capable of automatically classifying and measuring the length of fish observed in video. We will focus initially on underwater video, but much of the software will also be applicable to fish onboard commercial fishing vessels as well. Our approach is to first segment fish based on their motion in the water. For this purpose, we will make use of existing automated detection and tracking software developed at Toyon for the purposes of detecting and tracking people in video. In order to get a size estimate of the fish, we need to know the range to the fish and in order to get this information we propose the use of two camera and stereoscopic algorithms. For classifying the fish, we need to get a snapshot as close to broadside as possible for comparison with a library of feature models and we intend to select this image by maximizing the area of the segmented fish at a given range. At the end of Phase I, we will demonstrate the ability to automatically detect and track fish in underwater video and we will classify fish using a set of snapshots selected in cooperation with NOAA staff. We will create a confusion matrix for the classifier which quantifies the performance of the classifier when using the datasheet selected in Phase I. Finally, we will compare length measurements of fish made by an imaging system to physical measurements of the fish.