A neural network approach to rotated object recognition based on edge features: Recognition rate and CPU time improvement for rotated object recognition using DWT

In this paper, the methods and experimental results of rotated irregular shaped edge objects from still images and recognition by supervised artificial neural network of irregular shapes (ANN) are discussed. The edge images are obtained during pre-processing stage from the input coloured images. The objects in the edge images are rotated by every 5° from 0° up to 355°. The rotated edge images are represented by large dimensional feature vectors. A multilayer neural network classification system using backpropagation algorithm is trained to predict the classes of test images. Initially, features are extracted from rotated edge images. For the first part of the paper, the features are directly computed from rotated edge images and the artificial neural network (ANN) is trained. Training of ANN is done for 5 cases and in each case, images that are separated by varying angles of rotation are used. For the second part, the features extracted from the rotated edge images are subjected to 2D Discrete Wavelet Transform (DWT) to compress the dimensionality of features. The compressed features are then used for training the neural classifier. The five cases used in the previous part are considered in this part also. In both the cases, training is done with 75% of data sets and the performance of the neural classifiers is tested with the remaining 25% of the data sets. The CPU times, calculated during the training of neural networks in the two approaches are compared here.