Object Recognition Based On HDO Local Feature Description

It is a very difficult problem while detecting and recognizing simple object by machine or computer. One of the most critical problem is the representation of objects,i., what features will be favor to distinguish the difference between a specific object and other objects. Local invariant features is an effective way to solve this problem.Local features is an effective way to describe image local area and they present strong discriminative,stability, and localization ability. Therefore the local feature have been applied to solve much difficult computer vision problems. Because the local features is able to provide image content with statistical significance, and avoid the very difficult semantic level image segmentation. So it has great value in image understanding. Now it has become a mainstream technology in the field of object detection and recognition.The local features general require scale, brightness, rotation and translation invariance. HDO is a kind of local feature that is recently put forward to resist noise and illumination variance, and can be used in various object detection tasks, however traditional HDO do not have ability to resist rotation transformation, which limits its application. This paper studies Aluminum plate surface surface quality detection based on fusion HDO feature, and also improve traditional HDO feature with rotation invariance.(1) After image preprocessing and defect object segmentation, HDO feature and other related grayscale features, geometry features, texture features of the detected target area are extracted. All of these features are combined together to form the features of the defect image. And Adaboost-BP neural network is designed to recognize six class defects. The experimental results show that: after fusing HDO,the average classification recognition accuracy reached 90.3%, which is higher than traditional method(87.0%) when only grayscale, geometryand texture feature are adopted.(2)Because the traditional HDO local feature has the shortcoming that do not have rotation invariance, this paper presents a way to improve rotation robust with HDO feature description. After RGT(Radial Gradient Transform), the structure tensor of given location is calculated by circular neighborhood to acquire rotation invariant dominant orientation and coherent feature.Then, to enhance distinctivess, space pooling operation is implemented with Multi-Sector division. Test result in publicMIT human faces data shows that if the image doesn't rotate, the proposed method and the original HDO descriptor almost have the same accuracy(92.10%), while, if the image rotate, the proposed HDO descriptor is more accurate than the original one by 10.36%. In addition, in the experiments of synthetic rotated palms and faces recognition, the test result of our proposed HDO descriptor is obviously superior to its traditional one. The results show that the proposed HDO has strong ability to resist rotation transformation.