Object category recognition is a challenging task due to the low level and non-discrimination in visual representation.Most previous methods concentrate to find better high level visual features. Recently, optimally integrating various features to solve the problem attracted more interests. In this paper, we provide a novel method for object category recognition by improving the popular bag-of-words (BoW) methods from the following two aspects. First, we propose to extract a series of high level visual features which exploit both the local spatial co occurrence between low level visual words and the global spatial layout of the object parts. To obtain the global spatial features, a fast method is proposed to generate the semantic meaningful object parts by exploiting the geometric position distribution of the local salient regions. The image part patches are further quantized as semantic coherent high level visual words by using correlational spectral clustering. Based on it, simplified 2D string representation is introduced to model the global spatial patterns of the objects. Second, a multi-kernel learning framework is proposed to adaptively integrate extracted features in an optimal way. For each object class, an optimal feature weight coefficient is learned automatically and separately to combine both the low level and high level visual features by considering their contribution for the different object class. The tests on Caltech-101 and Pascal- VOC 06 dataset demonstrated that our method outperforms the baseline method BoW and state-of-the-art Multi-CM model [14].