Robust Classification Approaches to Industrial Sorting of Herring Fractions
Abstract
Among the rest raw material produced during the filleting process of herring there are high value products such as roe and milt. As of today there has been little or no major effort to process these by-products at an acceptable state, except for manual separation and mostly they are mixed into low-value products. Even though the roe and milt can be sold for as much as ten times the value of the mixed product, the separation costs using manual techniques render this economically unsustainable. The potential for better utilization of these products is large, as they contain high amounts of nutrition. Also, automating this extraction process could have the potential of giving the pelagic fish industry an additional income.
In this thesis, automatic separation of the by-products based on their distinct features is carefully analyzed. The analysis is conducted using new data from samples extracted from image recordings of by-products delivered by a processing factory. Different camera modules are tested in order to find the optimal solution for the imaging. Further, the data is divided into three respective classes; roe, milt and waste (other) before model tuning and analysis using multiclass support vector machines (SVMs), grid-search and cross-validation is applied to investigate the separability. Lastly, multiclass SVMs are implemented in LabVIEW for a realistic classification test in order to automatically predict the relationship of new incoming herring by-products.
Both the analysis and the results from the final testing indicated that the three mentioned classes are close to being nonlinearly separable using kernel SVMs, in this case using the radial basis function. More precisely, it was shown that good results were achievable considering the pairs milt/roe and roe/waste. Separation of milt and waste proved to be the most difficult task, but it was shown that acceptable trade-offs between accuracy and cost related to the classification could be achieved using certain model selection schemes. One of these schemes involved the maximization of the precision; the true positive rate of the predictions.