Training method of local description sub-network for natural scene image matching

The invention belongs to the field of image processing, and particularly relates to a training method of a local descriptor network for natural scene image matching, which comprises the following steps of: training a descriptor network by adopting an original sample and a generated sample; in each training, samples are generated by adopting the following modes: sampling each original sample in a geometric perturbation mode to generate each new sample; calculating descriptors of the new samples and the corresponding original samples by using the current descriptor network, calculating an adversarial generation loss function through the descriptors, and performing back propagation and optimization control parameters based on the adversarial generation loss function so as to generate a generated sample with a larger adversarial generation loss function; a loss function used for each iterative training comprises original sample three-input loss and generated sample three-input loss, the distance in each loss is c