Deep Learning Based Object Recognition from RGB-D Images
In the field of computer vision, image recognition has been developing for a long time. Text recognition, license plate recognition, etc. are already very mature technologies. In recent years, with the development of deep neural networks, the technology of object recognition has been further developed. Starting with RCNN(Region-CNN), although the accuracy of recognition is constantly improving. But at the same time, object recognition faces many challenges. These challenges include lighting conditions, similar colors for object and backgrounds, and more.
To meet these challenges, this thesis proposes a method to improve the accuracy of object recognition model by using depth information. This method uses the Grab-cut algorithm segmentation the depth information and uses the depth map after the segmentation to complete the segmentation of the target object. This method avoids the impact of complex scenes on object recognition. Using this method also reduces the effects of illumination, shadows, colors, etc. on recognition accuracy. The effectiveness of our proposed method is demonstrated by testing the depth map database we collected. As a result of the experiment, the average accuracy of the method can be improved by 5% to 10%.