Multispectral image co-occurrence matrix analysis for poultry carcasses inspection

Textural feature analysis of multispectral images containing visible/near-infrared (NIR) wavelengths based on co-occurrence matrices was demonstrated as feasible for discriminating abnormal from normal poultry carcasses at a wavelength of 542 nm. Statistical regression models and neural network models were used to develop classifiers. The results showed that the accuracy for the separation of normal carcasses was 94.4% when the statistical regression model was used. Specifically, the classification was perfect when the normal carcasses were separated from the septicemic and cadaver carcasses. However, for separating condemned carcasses between septicemic and cadaver, the accuracy was 96% for septicemic and 82.7% for cadaver cases. When neural network models were employed to classify poultry carcasses into three classes (normal, septicemic, and cadaver), the accuracy of classification were 88.9% for normal, 92% for septicemic, and 82.6% for cadaver cases. Whereas, the separation of the neural network classifier performed without error for two classes (normal vs. abnormal) classification.