An Unsupervised Feature Point Detection Network Based on Difference Response Graph

In order to break through the performance limitations of hand-craft feature point detectors, a new data-driven unsupervised feature point detection network based on differential feature response graphs is proposed. The network calculates the differential output using convolution kernels of different scales, uses the absolute value of the differential response map to select a large number of feature points and judges the importance of these feature points. The detector is trained by using various image transformations such as rotation, illumination, and blur to obtain corresponding feature invariance, making the network more suitable for small-scale data set training. Through experiments on three commonly used data sets, and qualitative and quantitative comparison analysis with existing classic algorithms, the experimental results show that the proposed unsupervised feature point detection network based on the differential response graph can complete the feature point detection well. The network locates object edges more accurately, requires less training time, and has less dependency on dataset size. The network has optimized and improved feature point detection performance.