Genetic Tuner for Image Classification with Robotic Applications

This paper introduces a genetic tuner implemented for optimizing the image scene analysis prior to the grasping of the objects, to realize a concrete working robot named COERSU. Firstly, different architectures of the adaptive neuro-fuzzy inference system, multi-layer perceptron and K-nearest neighborhood classifiers are compared to perform scene analysis and object recognition. Following on, the MLP classifier is chosen due to its accuracy and flexibility to be tuned by genetic algorithm. The real-time experiments (after tuning) show that the performance of the genetically tuned MLP classifier is improved in terms of accuracy due to this hybridization. Finally, snapshots of the experimental results from COERSU in a table-top scenario to manipulate some soft objects (e.g. fruit/egg) are provided to validate the methods.