Many previously proposed methods of object recognition use the salient regions of the objects to improve their robustness to distortion and occlusion. The methods based on salient regions inevitably encounter the difficulties if several different objects share identical or similar salient regions. Moreover, if the salient regions cannot be selected very carefully, the performance will be deteriorated incredibly. Therefore, in this dissertation, we propose a method which uses discriminative regions rather than salient regions to perform object recognition. Our proposed method consists of two phases, namely, training and testing. In the training phase, we first use sliding windows of different sizes to retrieve a number of regions from an object. For each region retrieved, we extract a feature vector, each of which contains four types of descriptors, namely, color histogram, intensity moments, affine invariant moments, and SIFT descriptor. Then, the Crisp Construction Process algorithm is applied to these training feature vectors to generate a number of prototypes for each model object. The prototypes of a model object can be used to discriminate it from the others. That is, the prototypes are the discriminative regions of the model object. In the testing phase, we also use sliding windows to extract the feature vectors of a test object. For each feature vector extracted, we find its nearest prototype and assign it to the discriminative region represented by the nearest prototype. Then, we compute a score for each model object according to the area of the model object, the area of the test object, the area covered by the feature vectors that are assigned to the model object, and the area covered by the assigned discriminative regions. The test object is considered as the model object with the highest score. Moreover, we adopt C4.5 decision tree to speed up the recognition process. Our proposed method is robust to distortion, occlusion, illumination changes, and cluttered background. Noisy and compressed images can also be well recognized. The experimental results show that our proposed method outperforms the comparing methods in the COIL-100 and ZuBuD datasets in terms of recognition rates. By adopting the C4.5 decision tree, the recognition process becomes 5 - 8 faster.