The Research And Application Of Local Invariant Feature Description Algorithm

Image local invariant feature is a research hotspot in the areas such as image processing, computer vision, pattern recognition. Since the performance of local image invariant feature is better than image pixel, the local image invariant feature has been widely used in many computer vision applications such as image matching, image stitching, image retrieval, object recognition and tracking.Image feature description is a main process of image local invariant feature detection which converts the image to a vector form in mathematics, so the image feature can be used in image processing. Local Intensity Order Pattern (LIOP) is a new image feature description algorithm. This algorithm design a local coordinate to make the descriptor invariant to image rotation. This method can avoid estimating the direction of image feature. However, the algorithm is vulnerable to noise when order the sampled pixels. And, the descriptor’s dimension is too high while we sample more pixels around the neighborhood, so the descriptor is not convenient to after-utilization.In view of the disadvantage of LIOP, we have made the following improvements:(1) We present an improved method of LIOP based on threshold to reduce the impact of noise. When we order the sampled points, there is a relationship of size between two sampled points while their gap is more than the given threshold. Thanks to this method of setting a threshold, the impact of noise is reduced effectively.(2) In view of the dimension is too high, we present a method of dividing the sampled points into two parts by their sequence number (odd or even), and construct descriptor by the method based on threshold respectively. Then, we connect two descriptors to get the final descriptor of this pixel. This method can reduce the dimension of descriptor while the performance is guaranteed, so the descriptor is convenient to used in the application of computer vision, such as image match.(3) The method of the weighted texture spectrum was used to cumulate the local descriptor. The weight is the summation of squares of the sampled points. This method further enhances the robustness of the descriptor.The descriptor has been evaluated on the standard Oxford dataset and some four additional image pairs with complex illumination changes. The experimental results show that our descriptor obtains a significant improvement over the traditional descriptors. We found that our descriptor is not only invariant to monotonic intensity changes and image rotation but also robust to many other geometric and photometric transformations. At last, we implement image stitching by our descriptor. From the experimental results we can see, if the scope of the image has enough size, the result is good enough for human’s visual requirements in the condition that we can ensure the image resolution is enough.