Hyperspectral detection of blood and melanin defects in salmon fillets

Summary

This study investigates the use of hyperspectral imaging (HSI) and neural network classification for detecting blood and melanin spots in salmon fillets. Hyperspectral images were collected and categorized into training (100 fillets), cross-validation (28 fillets), and test (229 fillets) groups to develop and evaluate a neural network classification model. Model classification performance was assessed at both pixel-level and at the object level by using a connected component analysis to detect connected regions in the classification maps. Blood spots were detected with higher recall (0.81 at pixel level and 0.92 at object level) than melanin spots (0.71 at pixel level and 0.86 at object level) in the holdout cross validation set, likely due to the distinctive absorption spectrum of hemoglobin compared to eumelanin. The corresponding F1 score, reflecting the harmonic mean of recall and precision, was 0.76 for blood and 0.68 for melanin at the pixel level. The F1 scores at the object level were higher, with scores of 0.85 for blood and 0.84 for melanin, since mismatch in the exact boundary of the defects affects the pixel level scores more than the object level scores. In general, the model effectively separated between blood and melanin defect types, although transient states between pure blood and pure melanin were also observed, particularly in the predicted class probabilities, in line with observations made by previous studies. Directly analyzing the modelled probabilities of blood and melanin may better resolve these transient states, rather than simply classifying the most probable class. The model maintained robust performance across different test sets, with an F1 score of 0.86 achieved for melanin spot detection in an image set (120 fillets) acquired at a different facility, indicating its ability to generalize to new data. Spatial registration of all images with a single reference fillet allowed a common spatial reference to be established. Spatial mapping of defect locations and analysis of trends is a promising means of gaining further insight into production conditions contributing to quality deviations. The contribution of the present study is largely a proof of concept, but the automated and rapid, non-invasive methodology employed is highly suited to the collection of big data sets that can contribute to better understanding the multifactorial development of blood and melanin defects. This study highlights the potential of HSI and neural network classification in enhancing quality control processes in salmon production.