Text-based target detection method in dynamic scene

Abstract

The invention discloses a text-based target detection method in a dynamic scene. Two memory mechanisms of implicit memory and explicit memory are constructed; the importance degree of the model is stored through implicit memory; and the training of the next batch is restrained through the importance degree, the data difficult to migrate is stored through explicit memory, and the data difficult tomigrate is put into the training set of the next batch to be trained, so that the neural network model can be effectively helped to memorize, and the problem of catastrophic forgetting of the neural network model in a dynamic scene is avoided.

Description

technical field [0001] The invention belongs to the field of text target detection, and in particular relates to a text-based target detection method in a dynamic scene. Background technique [0002] Since the success of deep convolutional neural networks on classification tasks in 2012, the research and development of computer vision has entered a golden age. Considering the further requirements of human-computer interaction, the research on combining computer vision and natural language processing has been on the rise in recent years. Compared with target detection, text-based target detection not only requires the computer to understand the category and location information of the image instance, but also requires the computer to use the internal attribute information and external relationship information of the instance to predict the image described by the text on the basis of understanding the text. instance. Since text-based object detection achieves semantic alignm...

AI Technical Summary

Problems solved by technology

However, there are the following problems in the existing technology: when new data arrives, the old data will not be visible to the model; the model needs to learn the knowledge contained in all batches of data, and directly applying the existing text-based target detection model to this task will be difficult. The classic catastrophic forgetting problem occurs. The earlier the data is learned, the more the model performance decreases and the greater the degree of forgetting

Images