A discussion on mass function generation for EORELM-based apple grading

Apple grading has become an important research topic in recent years. In order to improve the performance of automatic apple grading and sorting system, in this paper, evidential ordinal extreme learning machine is applied to model the relation between apple features and grades. Based on the ordinal extreme learning machine, evidence ordinal extreme learning machine can further deal with the epistemic uncertainty of grade labels modelled by mass functions. Considering the uncertainty of labels and the ranking relationship between labels of apple training example, the triangle mass function and Gaussian mass functions with different standard deviation σ values are designed and detailed. In addition, three evidential coding schemes are compared, each of which outputs a different coding vector with a given mass function. Experiments on Red Fuji apples show the grading performances of the single-model multi-output EORELM models with different mass function generation approaches combined with distinct coding schemes. Compared with traditional ELM model and ordinal ELM model, EORELM can obtain better grading performance.