Intelligent video image retrieval method based on neural network self-temperature fault and knowledge conduction mechanism

Abstract

The invention discloses an intelligent video image retrieval method based on a neural network self-temperature fault and a knowledge conduction mechanism, which improves the retrieval precision of a small model while ensuring the real-time performance of the model, and achieves balance between the precision and the efficiency as far as possible. A gamma correction module is arranged, and through local adjustment of the image, illumination non-uniformity robustness is achieved, detail discernibility is improved, high-frequency noise is avoided, and universality is high; a self-temperature fault mechanism is established, local self-supervision, continuous reflection and learning parameter adjustment of the neural network are allowed, deep semantic information of the image is fully learned, rapid convergence of the neural network is achieved, and retrieval precision is improved; a knowledge conduction mechanism is adopted, the model precision is improved, the model time delay is reduced, network parameters are compressed, and finally a student model with high performance and high precision is obtained; and taking shallow feature knowledge as a learning target through a conduction mechanism, reconstructing deep features by adopting a VAE variational self-encoding model so as to generate a learning result, and measuring the learning result and the target so as to complete a learning task.

Description

technical field [0001] The invention belongs to the field of image retrieval, and in particular relates to an intelligent video image retrieval method based on a neural network self-review and knowledge transfer mechanism. Background technique [0002] Image retrieval is the quantitative analysis of the image, so as to realize the correct judgment of the image. Some scholars have introduced it into the field of mineral processing and underground image retrieval. However, with the explosion of data volume, algorithms based on support vector machines and linear classification are limited by the ability of data representation, so the retrieval effect is not ideal. There are related studies that combine support vector machines with neural networks to achieve image prediction by replacing the softmax layer. However, support vector machines need to find suitable spatial parameters in high-dimensional data to generate an effective interval classification layer, thus increasing th...

AI Technical Summary

Problems solved by technology

[0009] In view of the problems existing in the above-mentioned background technology, the present invention proposes a kind of intelligent video image retrieval method based on neural network self-warming and knowledge conduction mechanism, and for the uneven and complex environment of underground illumination, the present invention proposes a deep gammacorrection (gamma correction) module, By learning the knowledge of the local light and shade distribution of the image, the visibility of details can be improved and the robustness to uneven illumination can be achieved; in view of the common problem of poor real-time performance of deep learning algorithms in the mine, a lightweight network is adopted to ensure low time cost In this case, the efficient retrieval of foreign object images can be realized; a self-review and knowledge transmission mechanism is proposed, and the robustness of lightweight networks can be improved by teaching knowledge through large models

Images