Research On Infrared Polarization Image Processing And Data Enhancement Method For Ships In Sea Surface Background

Maritime infrared ship target detection is of great importance in military and civilian applications;however,vulnerable to environmental and camouflage target interference in target detection,resulting in low detection accuracy.Secondly,infrared ship images are difficult and costly to acquire,and cannot provide sufficient training data sets for target detection algorithms.In this thesis,we enhance the weak targets of sea surface ships by fusing ship infrared images with polarized images,and use adversarial neural networks to enhance the image data,to provide sufficient training data for target detection algorithms.The research of this thesis mainly includes the following aspects:1.To address the problem of low detection accuracy due to unclear targets of ships at sea,this thesis proposes a fusion method of ship IR and polarization images based on VGG network,decomposes the IR and polarization images into contour part and detail feature part,fuses the contour part by using weighted average strategy,and the detail feature part is deeply extracted by VGG network and reconstructed by multi-layer fusion strategy,and finally The fused image is obtained by combining the contour part and the detail feature part.Through the comparison of contrast and signal-to-noise ratio and other indexes,the performance of the proposed algorithm in this thesis is significantly improved compared with the traditional fusion method,which enhances the weak targets of ships on the sea surface.2.To address the problems that the current ship target detection and recognition algorithm lacks training data set,and it is difficult and costly to obtain infrared images of ships at sea,this thesis proposes an image data enhancement method based on single-image generation adversarial network(Sin GAN),in which the model is trained by learning a single ship infrared polarization fusion image,and the trained model gets multiple generated images with different morphological structures by a random noise input.images with different morphological structures.To address the problem of confusing ship target structure and poor connectivity in some types of ship data enhancement results,this thesis proposes a parallel Sin GAN adversarial generative network to train multiple generators at the same time,using high learning rate for the generator training of low-resolution images and low learning rate for the generator training of high-resolution images to obtain high-quality and structural integrity of the generated images of ships.3.For the problem that the generated images have missing details and low resolution,the generator network and discriminator network are improved by using the anti-bottleneck structure of Conv Ne Xt network,using large convolution kernel to improve the network accuracy by group convolution,combining with the residual structure of the model itself to supplement the missing information,and up-sampling to get the high-resolution output images.In summary,this thesis improves the performance indexes such as the contrast of images by fusing the infrared images of ships in the sea surface background and polarized images,and generates a large number of high quality images of ship targets on the sea surface by super-resolution reconstruction of single-image generation adversarial network,which solves the problem of small ship target detection samples and difficulty in supporting training,and provides help for subsequent detection and recognition.