RGB-T image salient target detection method combining independent decoding and joint decoding

Abstract

The invention discloses an RGB-T image salient target detection method, which comprises the following steps: extracting RGB features and Thermal features from an RGB-T image, generating a corresponding initial saliency map, enhancing the RGB features and the Thermal features by using the initial saliency map, implementing independent decoding, and generating RGB and Thermal independent decoding features and an independent decoding saliency map; fusing the RGB independent decoding features and the Thermal independent decoding features to form RGB-T fusion features, and then implementing joint decoding to generate a joint decoding saliency map; and supervising the independent decoding saliency map and the joint decoding saliency map by using the saliency map truth value, forming an RGB-T image saliency target detection model through training of the training data set, detecting any RGB-T image by the model, and outputting the joint decoding saliency map. According to the method, the uniqueness of the RGB color image and the uniqueness of the Thermal thermal infrared image are considered through independent decoding, the difference between the RGB color image and the Thermal thermal infrared image is reduced through joint coding, and the detection performance is improved.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to an RGB-T image salient target detection method combining independent decoding and joint decoding. Background technique [0002] RGB-T images are composed of RGB color images and Thermal thermal infrared images. When the light of the RGB color image is insufficient or the background noise is too much, the effect of image salient object detection is often unsatisfactory. However, thermal infrared cameras can capture infrared radiation emitted by any object whose temperature exceeds absolute zero, so Thermal thermal infrared images are not sensitive to lighting conditions and have strong anti-noise interference capabilities. Combining the advantages of RGB color images and thermal infrared images, salient target detection in RGB-T images has gradually become a research hotspot. [0003] Transformer relies on multi-head self-attention and forward propagation networks to make breakthr...

AI Technical Summary

Problems solved by technology

[0004] The key to salient target detection in RGB-T images is to effectively mine the information of the two modes of color images and thermal infrared images and realize their fusion. Dual-stream encoding and joint decoding are the main methods used by the existing RGB-T image salient target detection models. , it tends to consider the fusion of RGB color images and Thermal thermal infrared images, and does not consider the uniqueness of RGB color images and Thermal thermal infrared images

Images