Deep-learning classification of chicken woody breast based on bioelectrical impedance characteristics

Abstract

As a serious threat to the broiler industry, woody breast (WB) requires precise classification that is theoretically aligned with the advantage of bioelectrical impedance detection. This research used normal chicken breast (NORM) and three levels of WB condition, namely, mild, moderate and severe (SEV), based on sensory evaluation. The basic objective quality indicators and impedance characteristics of the samples were detected, and then the various levels of WB were categorized by model-classification approach. At a consistent frequency, the impedance amplitude of samples decreased with increased WB level. Significant differences in the absolute value of the phase angle existed among different levels of WB. The increase in WB level led to a considerable increase in intracellular resistance (Ri) and in the characteristic frequency (fc). However, four other indices including the radius of Cole-Cole curve arc, the extracellular resistance (Re), the polarization coefficient (K), and the relaxation factor (α) substantially dropped with increased WB level. The accuracy of SEV training, NORM and SEV test samples achieved a perfect score of 100% according to the partial least squares (PLS) prediction model. The PLS model also exhibited an overall accuracy of 91.70% for training samples compared with the value of 88.35% from limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) deep-learning prediction model. However, the L-BFGS model achieved a higher overall correct rate for test samples (90.00%) than PLS model (80.00%). These results provided valuable information for the classification of WB based on the characteristics of bioelectrical impedance.

Keywords

• woody breast
• bioelectrical impedance
• partial least squares
• limited-memory broyden-fletcher-goldfarb-shanno
• discriminant classification