An appearance based neural image processing algorithm for 3-D object recognition

We propose an appearance based neural image processing algorithm for the recognition of 3-D objects with arbitrary pose in a 2-D image. Instead of object segmentation we utilize the wavelet transform to extract compact features for object representation. Translational invariance is achieved by using two neural network based object pose estimators to translate objects automatically to the image center. Based on these translation-invariant features a neural model is built to identify objects taken at different viewpoint and under different illumination condition. Results for the recognition of real images under occlusions are shown