Derinlik kamerası görüntülerinden nesne tanıma

Object recognition from RGB-D images that provide additional depth information is very important task in many real world robotics and computer vision applications. The Convolutional Neural Networks (CNNs) have widely used in numerous applications especially RGB-D object recognition. However, CNNs have several restrictions even though they have demonstrated outstanding performance on object recognition. Pooling layer of CNNs causes to information loss in the stage of feature extraction. In addition to this, CNN is very sensitive to environmental factors such as rotation and light intensity. Capsule networks proposed by Hinton have been developed to avoid from these problems. In the thesis, the performances of the Capsule networks are investigated on the RGB-D dataset. Also a two-layer hierarchical structure is proposed in which the depth images are used in the first layer and RGB images are used in the second layer. Two different hierarchical structures that consist of CNN and capsule networks are designed. The performances of the hierarchical CNN and capsule networks are evaluated on the Washington RGB-D dataset. According to the simulation results, the best performance has been achieved with hierarchical CNN.