Research On Object Recognition Based On Bags Of Words

Object recognition is one of the most active research domains in computer vision currently.It has good application prospects, for example, growing surge in military and civilian requirements, intelligent navigation, large scale of content-based image retrieval in Internet, escort robot, automatic image labeling and so on. The ability to identify objects is one of the basic functions to meet these requirements.With the constant progress of artificial intelligence and pattern recognition technology, the existing object recognition technology has achieved good results, but there are still some problems. At present, the methods based on the underlying global image features are unable to overcome the interference from background, occlusions, scale and angle change. Besides, the algorithm based on components and structure tends to ignore the location information between the parts in images.The object recognition method based on local features of image have the characteristics of low computational load and high content of information. Specifically, the scale invariant feature transform (SIFT) descriptor is widely used in object recognition. But the number of SIFT feature points extracted from image is usually uncertain and it cannot be inputed directly into the classifier for the trainning. Conversely, Bags of Words model which is based on local features can represent the sets of local image features as a fixed dimension vectors and the vectors can used as the inputs of machine learning algorithm for object recognition.In recent years, the Bags of Words model based on local features used in the scene and object recognition has made brilliant achievements. However, the basic Bags of Words Model cant’t get the shape or segmentation of the object from the background, because it can’t distinct background and foreground when it covers a whole image. In addition, the model completely ignores spatial information of the local features of images and has some limitation in the description.After thoroughly analyzing the difficulties of object recognition and considering the advantages of local features and Bags of Words model, this dissertation researches the methods of object recognition based on Bags of Words model and do some improvement to solve the problems existing in the traditional Bags of Words model. Experimental results proved the effectiveness of the method proposed in this dissertation. The main research contents and innovation points are as follows:(1) In order to solve the problem that many local features are unstable, unreliable or irrelevant in constructing visual dictionaries, a new object recognition method is proposed based on Salient Regions and BOW Model. In this dissertation, we locate the regions of interest in images by using local invariant feature point detection operator and according to the distribution of invariant feature points instead of utilizing the complex image segmentation technology. Local features are extracted in the region of interest, which can effectively reduce the feature points have nothing to do with the object, ensure the extracted features can describe the features of the images more accurately and reflect the nature information of the object, also can resist the impact of the various features from the background.(2) Inspired by the spatial pyramid model, an optimized method is proposed based on multiple directions spatial Bags of Words model for object recognition. In order to take full use of spatial information between image parts, projection method is applied to local image blocks to get spatial structure information of images. By combination of sample codebook, the final representation of features which are extracted from the image abounds more vision sense.The experiments are carried out based on public object recognition database, the results show that the optimized methods can achieve better performance compared to the traditional Bags of Words method.