Automated Filet Identification

In the United States and other countries where seafood is often sold in the filet form there is an incredible amount of illegal mislabeling of market species for economic gain. This is made possible because the average consumer cannot identify fish species in the marketplace, especially in filet form. One seafood fraud study by Oceana reported the mislabeling of seafood at 33% in the United States. To solve this problem, we will develop a species identification methodology based on visible imagery that relies on commodity hardware such as tablets and smartphones. The algorithm, deployed in a user-friendly mobile application, will use efficient deep neural networks trained on a large amount of filet imagery started during phase I and expanded in phase II in collaboration with two of the largest seafood companies in New England: Blue Harvest Fisheries and Gulf of Maine Sashimi. Our expanded dataset will include more fish species and more images per species. Building off encouraging performance achieved in phase I, the algorithm will continue to be improved both in terms of classification accuracy and model calibration which will give users of the application an understanding of how confident the algorithm is in any decision that is made.