Supervised learning models for control quality by using color descriptors: A study case

This paper presents a study case for color inspection in quality control applications using color descriptors histogram RGB-1D and histogram TSL and supervised machine learning methods such Support Vector Machine (SVM) and Artificial Neural Networks ( ANN). For this, we build three annotated databases, and these are made using real application of quality control like color inspection in forages and polarized level from vehicle glasses. These bases are captured with a Samsung Galaxy S5 mini camera, which has a resolution of 800x480 pixels. Each class has fifty images under uncontrolled conditions of noise and lighting. The first databases consists of living colors required for Wood fodder with texture. For the third one, it takes glasses with different level of polarized. To calculate the learning methods performance, we use a cross-validation method, which fractionates the data (70% for training and 30% for validation). In the ANN test setup, we use a Backpropagation algorithm. For the SVM case, we take a multi-class setup with Gaussian Radial Kernel (RBF) that uses an adaptative radio with classification strategy "one-vs-all". Finally, it is reported the accuracy average for each class and its standard deviation.