Multi-spectral image analysis using neural network algorithm for inspection of poultry carcasses

A multi-spectral imaging technique was implemented in an on-line inspection system for the separation of wholesome and unwholesome chicken carcasses. A multi-spectral imaging system provided image information of the carcasses in the spectral and the frequency domains. The system acquires spectral images from the chicken on a moving shackle in real-time and processes these spectral data for classification. The spectral images of 540 and 700 nm wavelengths were useful for separating unwholesome carcasses having characteristics such as ascites, air sacculitis, bruise, cadaver, leukosis, septicemia and tumor from the wholesome carcasses based on spectral image pixel intensity and the intensity distribution of Fourier power spectra. The mean intensity of the wholesome carcasses scanned at 540 nm wavelength was higher (P less than or equal to 0.01) than the intensity of unwholesome carcasses. On the other hand, when Fourier spectrum pixel intensity at 700 nm wavelength was used, the intensity of wholesome carcasses was much lower (P less than or equal to 0.01) than unwholesome carcasses. The accuracies of neural classifiers were 100% for calibration and 93.3% for validation when combined spectral image pixel intensities of 540 and 700 nm wavelengths were used as inputs. The accuracies were 93.4% for calibration and 90% for validation when fast Fourier transforms of image intensity data of 700 nm wavelength were used as inputs for a neural network model. (C) 1998 Silsoe Research Institute.