Research On Object Recognition Based On Interest Point And Fusing Multi-features

Object recognition is one of the most challenging researches in the field of computer vision. With the improvement of modern image capturing technology, object recognition based on image has nicer foreground. However, there are many difficulties in object recognition based on image, such as viewpoint changing, difference in inter-classes or intra-classes, which make the object recognition challenging.In recent years, interest point is widely used in image recognition because of its advantages such as small computing and rich information. Most of the methods of object recognition use single feature. But the single feature is not adapted to multi-classes object recognition, the result of recognition is not well; the reason is that we have to consider multi-features for the sake of improving the recognition rate of multi-classes of objects. Therefore, fusing multi-features was more and more used in pattern recognition. This paper considered advantages of interest point and fusing multi-features, proposed an object recognition method based on interest point and multi-features.Firstly, we use simplified Local Ternary Patterns to wipe off redundant Harris corner. We compared the advantages of Local Binary Patterns and Local Ternary Patterns and found that Local Ternary Patterns is more accurate. So, we use simplified Local Ternary Patterns to wipe off redundant Harris corner. The experimental results show that simplified Local Ternary Patterns is very effective for wiping off redundant Harris corner.Secondly, we defined a method to find an interest area of image according to the location of corner. The experimental results show that the definition of interest area is very benefit for extracting feature because of that the interest area include main information of object in the origin image and wiped out some redundant information such as background.Thirdly, we extract shape, texture and color feature of interest area by using shape invariant moment, discrete wavelet transform and color histogram respectively. This method can not only robust to translation and rotation, but also overcomes the drawbacks of losing the location information in three features. Finally, we give a method of fusing multi-features combined with Support Vector Machine and K-Nearest Neighbor. We use Support Vector Machine to classify object according to single feature. And we compute weight of three features according to the results of classifying. Most importantly, we introduce weight of three features to K-Nearest Neighbor and improve distance function. In this way, we have achieved a classifier adapting to every class of object. We make experiments on image library--Caltech-101. Experimental results show that the improved method effectively improves the accuracy of object recognition.