A comparative experimental study on the prediction of renewable oils properties using RGB and HSV image processing techniques

Abstract

In this study, renewable oil properties of Flash Point (0C), Fire Point (0C), Density (kg/m3), Cloud Point (0C), Pour Point (0C), and Viscosity (cST) are predicted using image processing techniques of Red Green Blue (RGB) and Hue Saturation Value (HSV). Eleven types of renewable oils are chosen for experimentation, and their surface images are captured with a high-resolution digital camera. For better accuracy, around 150 surface images are captured for each oil sample, and their average pixel data is extracted using RGB and HSV techniques. The digital pixel information (metadata) of all the oil samples is mapped to their experimental oil properties, and the accuracy of the developed metadata is validated with Fiji software due to its better image analysis and also complex data quantifying capabilities. The minimum, maximum, mean, mode and standard deviation results of RGB and HSV agree with Fiji. In addition, the developed dataset has been validated with Neural Network classification and TreeBagger algorithms. The results of TreeBagger reveal that the trained dataset is highly accurate (91.9% for RGB and 95.3% for HSV). Similarly, 95.6% (RGB) and 97.3% (HSV) accuracy is achieved for Neural Network classification. Finally, two new oil surface images are trained using the developed dataset. Both RGB and HSV accurately predict the oil properties. Therefore, it is evident that predicting the significant oil properties helps optimize the production process by reducing experimental costs and time.