Random forest classifier and neural network for fraction identification of refuse-derived fuel images

The current paper aims at the evaluation of computer vision methods to identify RDF (Refuse-derived fuels) fractions based on images of the RDF particles. For this purpose, images of 1345 single RDF particles, with a typical size in the cm-range, were taken in bird’s eye view and assigned manually to one of six predefined material fractions. Two possible methods were tested for fraction identification: First, a classical machine learning approach consisting of feature extraction of color histograms and Haralick-textures as input for a random forest classifier and, second, a neural network approach with transfer learning. In the machine learning approach, the random forest classifier with feature extraction based on image color distribution (histograms of colors in images) and Haralick-textures (matrix of gray-value co-occurrences) achieved an accuracy of 49 %. The neural network approach is based on the Xception network, a state-of-the-art convolutional neural network with depthwise separable convolutions. The accuracy of the Xception network for recognizing RDF fractions with transfer learning is 69%, thus considerably better than with the machine learning approach and a good starting point for identifying the very inhomogeneous appearance of RDF fractions. The Xception approach was then used in a more realistic setup where the RDF particles are transported by a conveyor belt which allowed for simplified image acquisition. Here, some additional measures are needed for image recognition like edge detection performed by the Canny algorithm. The accuracy for a test data set could be increased to 71 % after using images from this setup to improve the neural network.