The present invention is applicable to the technical field of
image registration, and provides a method for constructing a dual-mode
deep learning descriptor based on
graphics primitives. In this method, the attribute category of the patch image is learned by labeling samples, and the geometry of the patch image is learned by using the
graphics primitive. Features, the attribute category and geometric features are fused to obtain the
feature vector of the local patch image, that is, the descriptor based on the
graphics primitive. The registration between patches is completed through the similarity of descriptor vectors, and the classification and description based on
machine learning descriptors are realized. The dual-mode
deep learning descriptor construction method based on graphics primitives proposed by the present invention is aimed at the classic
image registration method CPU. Disadvantages of a large amount of calculation, explore the descriptor classification method of GPU (
image processor) calculation. It mainly establishes a descriptor
training set, builds a multi-mode convolutional network, trains categories and geometric
modes on the GPU, and realizes classification and registration of local patch images. Solve the classification description method of the descriptor and the implementation problem on the GPU.