Infrared image target detection method and device based on difficult sample transfer learning

Abstract

The invention relates to an infrared image target detection method and device based on difficult sample transfer learning, belongs to the technical field of image processing, and solves the problem of how to utilize rich natural image data to assist in improving the performance of infrared image target detection. The method comprises the following steps: acquiring a target detection data set and dividing the data set into a training set and a test set, wherein the data set comprises a visible light data set and a corresponding infrared data set; the construction of the difficult sample transfer learning infrared image target detection network comprises the following steps: introducing a domain attention module and a path aggregation network module into a ResNet50 network to form an improved feature extraction network; performing different-source image transfer learning on the infrared image target detection network by using the training set to generate a target detection model; and inputting the to-be-detected visible light image and the to-be-detected infrared image in the test set into the target detection model to obtain a target detection result. Different domains activate different data domains through domain attention and predict targets of different sizes on different scales.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an infrared image target detection method and device based on transfer learning of difficult samples. Background technique [0002] UAVs use information perception ability to accurately identify targets (vehicles, ships, low-flying aircraft, etc.) in complex ground and sea backgrounds. [0003] The reason for the emergence of difficult samples is often due to the small number of samples. The difficult samples brought by sample imbalance cause the gradient update brought by this sample to be covered by a large number of simple samples during model learning, which reduces the recognition probability of difficult samples. Heterogeneous image migration automatic target recognition technology plays an important role in identifying friend or foe, sea situation monitoring, long-distance air-to-ground reconnaissance and surveillance, precision guidance, and combat management. It...

AI Technical Summary

Problems solved by technology

[0007] In order to adapt to the changing environment of the battlefield and the attack of enemy targets that are difficult to capture, it is very necessary to detect possible threats in advance, especially at night and the detection of moving targets is relatively difficult. Carrying out infrared image target detection is helpful. Natural advantages, but the existing infrared data is very small, it is difficult for general machine learning algorithms to get better results, and the cost of image labeling is also high

In the case of relatively scarce sources of infrared image data, only a small amount of infrared images are used for training or fine-tuning from scratch, and the network is very easy to overfit

Images