Hyperspectral image detection of pork freshness based on Active Learning

Abstract

<abstract> <bold>Abstract. </bold>Pork freshness seriously influence the nutrition, food safety and quality of life, research shows it can be reflected well by detecting the TVB-N (Volatile Base Nitrogen). But it is costly and time-consuming to measure the TVB-N of pork. As a method of machine learning, Active Leaning is well solve the deficiency of labeling samples in small-scale samples modeling. This article describes detailedly Active Leaning sampling algorithm which is used to select representative training samples to reduce the dependence of model on the measured value. In this paper, the mean of spectrum and image entropy of hyperspectral images are extracted to build model with TVB-N. For reflecting the performance of Active Leaning algorithm, there are two cases to be considered, in the first case, randomly select 135 samples from 180 samples as modeling samples at first, and the rest of 45 samples are set as testing samples, then respectively use Random, KS (Kennard and Stone ) and Active Leaning sampling algorithm to select n(n=15, 30, 45, 60, 75, 90, 105, 120, 135) training samples from modeling samples ; in the second case, respectively use Random, KS and Active Leaning sampling algorithm to select n training samples from all the samples, and the rest of samples are set as testing samples set. At last, build prediction model by LSSVM (Least Square Support Vector Machine) in two cases, and use RP (Relevance of Prediction Set) and RMSEP (Root Mean Square Error of Prediction Set) to comprehensively evaluate the performance of the model. when build model with mean of spectrum and the number of training samples is 15,in the fist case, the RP respectively are 0.43,0.51,0.59, and RMSEP respectively are 8.02,6.84,5.82 (Random sampling,KS,Active Leaning); in the second case, the RP respectively are 0.41,0.69,0.77, and RMSEP respectively are 6.40,5.45,4.79 (Random sampling,KS,Active Leaning).The results show that Active Leaning is remarkably better than the KS and Random sampling in two cases.