Aiming to solve the problems of false detection, missed detection, and insufficient detection ability of infrared vehicle images, an infrared vehicle target detection algorithm based on the improved YOLOv5 is proposed. The article analyzes the image characteristics of infrared vehicle detection, and then discusses the improved YOLOv5 algorithm in detail. The algorithm uses the DenseBlock module to increase the ability of shallow feature extraction. The Ghost convolution layer is used to replace the ordinary convolution layer, which increases the redundant feature graph based on linear calculation, improves the network feature extraction ability, and increases the amount of information from the original image. The detection accuracy of the whole network is enhanced by adding a channel attention mechanism and modifying loss function. Finally, the improved performance and comprehensive improved performance of each module are compared with common algorithms. Experimental results show that the detection accuracy of the DenseBlock and EIOU module added alone are improved by 2.5% and 3% compared with the original YOLOv5 algorithm, respectively, and the addition of the Ghost convolution module and SE module alone does not increase significantly. By using the EIOU module as the loss function, the three modules of DenseBlock, Ghost convolution and SE Layer are added to the YOLOv5 algorithm for comparative analysis, of which the combination of DenseBlock and Ghost convolution has the best effect. When adding three modules at the same time, the mAP fluctuation is smaller, which can reach 73.1%, which is 4.6% higher than the original YOLOv5 algorithm.

In distance-selected imaging, the contrast of laser images is reduced due to long imaging distances, insufficient laser power, and atmospheric turbulence. An enhancement algorithm based on the EnlightenGAN network is proposed to improve the contrast of laser images. Firstly, the laser images are acquired using a distance selection pass system to establish the laser image dataset and expand the dataset, and the traditional algorithm is used to enhance the images and establish the mapping relationship between low-quality images and high-quality images. The global discriminator based on PatchGAN with the improved VGG model is used to regularize the self-feature retention loss and construct the depth link between the global discriminator and the local discriminator to improve the generalization ability of the model; adjust the attention map to the second layer before the CLB convolution module and also add the residual structure in the second layer CLB to improve the robustness of the model; adopt the idea of gray-scale layering with a low drop and high rise to improve the self regularization mechanism to achieve the enhancement of the key region; finally, use the improved EnlightenGAN to fit the relationship between a low-quality image and high-quality image. Finally, EnlightenGAN is used to fit the relationship between low-quality images and high-quality images, extract laser image features, and enhance low-quality images. The experimental results show that the improved algorithm improves PSNR by 12.3% and 0.7% on average, SSIM by 57% and 10.3% on average, and NIQE by 21% and 13% on average compared to other algorithms and the original EnlightenGAN algorithm, respectively. The algorithm improves the signal-to-noise ratio and contrast of laser images with richer image details. It provides a new idea for pre-processing laser images.