Beef colour measurement with an RGB camera.

Spectrocolorimeters are the most commonly used instruments for measuring meat colour because they provide readings in the device independent CIELÃaÃbà colour space which corre- sponds well with the colour perception by humans. These instruments, that can accurately measure colour surface of a sample uniform and rather small, cannot completely represent the surface characteristics when it is not uniform and highly textured, as in case of the meat. In recent years, digital cam- eras have been used to measure meat colour since they pro- vide some advantages over a conventional colorimeter, namely, the possibility of analysing the entire surface of meat. As the digital cameras use the RGB colour model, which is a device dependent colour space, it is necessary to derive a conversion matrix to estimate CIELÃaÃbà parame- ters from RGB measurements. Eighty longissimus thoracis steaks were cut into 5Â5Â4 cm samples. After blooming, the samples were photographed using a NIKON 3100 digital cam- era. All the images were captured under controlled conditions with a reference grayscale target, used to correct white bal- ance. The RGB colours of meat were measured from JPEG images using ImageJ software. An area of 3.0 cm2 at the centre of each sample was evaluated. Immediately after, the true colour of each sample was measured by a MINOLTA CM- 600d spectrocolorimeter and the results were expressed in the CIELÃaÃbà space model. The conversion of RGB colour values to LÃaÃbà values was carried out using a quadratic model which considers the influence of the square of R, G and B variables on the estimation of LÃaÃbà values. R- square, root mean square error and the mean normalized error were used for measurement of differences between the values achieved with ImageJ and the spectrocolorimeter. Moreover, the predicted colorimeter colour coordinates from digital images were compared to real colorimeter values using CIE colour difference equation (DEÃ). The real and estimated values showed coefficients of determination of 0.67, 0.41 and 0.33 for LÃ, aÃ, and bÃ, respectively. The root mean square error between ImageJ and spectrocolorimeter was 1.48, 1.11 and 0.80 for LÃ, aà and bà values, respectively. The model showed an error of 1.16% and a standard deviation of 0.92. The DEà was equal to 2.5. The proposed method achieves a promising performance, however the acquisition of the images need some adjustments to improve the accuracy of the model.