A data self-distillation method based on batch processing knowledge ensemble (BAKE)

Abstract

The invention discloses the technical field of target detection, and particularly relates to a data self-distillation method based on batch processing knowledge ensemble (BAKE), which comprises the following steps: S1, giving a small batch of samples and a network in training, and storing similar samples; s2, avoiding self-cognition enhancement, and normalizing the affinity matrix A; s3, predicting samples in one batch, and then obtaining and forming a soft learning target; s4, realizing complete integration of knowledge. Compared with an existing knowledge integration method, the problem that traditional knowledge integration needs to be processed through an additional network or branch is solved, the needed calculation amount and memory overhead are saved, BAKE is applied to a network architecture, the classification performance of the network architecture can be improved, and wrong supervision caused in the processing process is avoided.

Description

technical field [0001] The invention relates to the technical field of target detection, in particular to a data self-distillation method based on batch knowledge integration (BAKE). Background technique [0002] In target detection, better target generation will lead to better detection accuracy and classification performance of the network, where the integration of knowledge has been proven to generate better soft targets, but the knowledge integration of multiple networks will bring Large amount of calculation and memory overhead, in the absence of knowledge integration, due to the limited amount of information carried by the network, good targets cannot be generated, thereby limiting the performance of the network, so it is necessary to develop a good knowledge integration data Self-distillation method. [0003] However, due to technical reasons, most technicians nowadays generally adopt the integration method of multiple networks or multiple branches. However, there ar...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to provide a data self-distillation method based on batch processing knowledge integration (BAKE), so as to solve the problem that different types of samples cannot be fully considered because technicians rely on additional networks or branches. thus increasing the cost of computation and memory as well as leading to erroneous supervision

Images