Automatic Inspection of Cod (Gadus Morhua L.) Fillets by Hyperspectral Imaging

Abstract

The manual trimming and inspection of cod fillets by candling, is considered the bottleneck of cod fillet processing. The operation is both labour intensive and expensive, reported to account for as much as 50 % of the cost with cod fillet production. Due to the high labour costs in Norway, it is of great interest for the industry to optimize this process. In this work a hyperspectral imaging system has been developed, capable of inspecting cod fillets, with or without skin, at a conveyor belt speed of 400 mm/seconds, corresponding to the industrial processing speed of one fish per second. The system is designed as proof of concept, and the algorithms are not implemented to be run in real time. A method for segmenting a cod fillet image into the respective parts: loin, belly and tail, using the centerline as a reference system, has been developed. The method is useful for selecting standardized measurement regions on the fillet, and used for extracting data for automatic freshness assessment. Freshness, as days on ice, can be predicted using spectroscopy in part of the visible region (450-700 nm). This can be done with an accuracy comparable to what is reported for sensory evaluation using a panel of trained evaluators. The same system is used for detecting fillets which have been previously frozen, both as whole fish and as fillets with skin. The results show a complete separation between the fresh and frozen-thawed samples. Similar mechanisms are affecting the spectra from fish stored fresh on ice, and fish that has been through the freeze-thaw cycle. The main variations seen in the spectra from cod fillets stored on ice, or frozen and then thawed, are due to oxidation of heme proteins in the muscle. This is supported by independent measurements using two different instruments, and by previous studies pointing to the visible region as the best region for freshness prediction. Detectingobjectsembeddedintissue, usingvisiblelight, isdifficultduetovariability in the optical properties of the surrounding tissue. A method for calibrating the spectral signature from small objects embedded in translucent material has been developed. This method uses the estimated local background spectrum to calibrate the hyperspectral image, and the method is evaluated for automatic nematode detection, using the hyperspectral imaging system, at a commercial cod fillet processing plant. The local calibration method is superior to using traditional spectroscopic pre-treatment methods, and reduces both spatial and spectral variations across the image. The results from the industrial test show that the system can detect nematodes in cod fillets with a performance which is comparable or better, to what is reported by manual inspection.