基于改进生成对抗网络的红外图像增强算法

Infrared image enhancement algorithm based on improved generative adversarial networks

摘要:

针对红外图像细节分辨率不高、目标边缘模糊而导致目标识别率低，虚警率高等问题，提出一种有监督的生成对抗网络结合胶囊网络的图像增强方法，以提高图像清晰度。首先，在生成器中，将Unet与全局上下文模块相结合，捕获丰富上下文信息，提高图像细节特征；在鉴别器中，将胶囊网络与Res2net结合，提取图像特征和结构，提高细节提取能力，增强生成对抗的模仿能力。实验表明，该方法能够根据图像内容，对图像所需区域进行细节重点增强，从而极大程度提高图像分辨率和视觉效果。

Abstract:

Aiming at the problems of low resolution of infrared image details, low target recognition rate due to blurred target edges and high false alarm rate, a supervised generative adversarial network combined with capsule network is proposed as an image enhancement method to improve image clarity. First, in the generator, Unet is combined with the global context module to capture rich contextual information and improve image detail features; in the discriminator, the capsule network is combined with Res2net to extract image features and structure, improve detail extraction and enhance the imitation capability of generative adversarial. Experiments show that the method is capable of detail-focused enhancement of desired regions of an image according to its content, thus greatly improving image resolution and visual effects.