Research On Image Super-resolution Algorithm For Object Detection

Today,with the rapid development of science and technology,the speed of information dissemination continues to increase.As the speed of information dissemination increases,people’s demand for information is becoming more and more stringent.From ancient times to the present,one of the best carriers for disseminating information is images.With today’s high-speed communication networks,people can send and receive higherresolution images to each other.In the field of computer vision,image super-resolution algorithms also make it easier for people to obtain high-resolution images.However,in some professional fields such as aerospace,satellite remote sensing,and medicine,researchers not only require images with ultra-high resolution,but also require more detailed information on the objects of interest contained in the images to assist related detection tasks.However,most of the current methods focus on improving the overall quality of the image,and the processing results for specific objects contained in the image,especially small-sized objects,are not very satisfactory.Therefore,when these algorithms are used to detect scenes,the achieved performance is often not up to the mark.With the development of deep learning,the emergence of generative adversarial networks provides researchers with a new idea.The super-resolution algorithm based on the generative confrontation network has achieved excellent performance.This method learns the characteristics of the original high-resolution image and generates a high-resolution image with more details.This thesis studies the image super-resolution algorithm based on generative confrontation network in recent years,and proposes an improved super-resolution algorithm framework for detection scenes.By comparing the generative confrontation network proposed in recent years for super-resolution reconstruction,this thesis chooses to use real-ESRGAN as the basic network for research,because it has more detailed information about generated samples than other networks,and has adjustable magnification the advantages.Considering that the target of interest in the detection scene accounts for a relatively small proportion of the overall image,in order to avoid some of its features being ignored due to the down-sampling operation of feature extraction,this thesis proposes a scheme to introduce an attention mechanism in the network to provide a small target Assign more weight.This thesis chooses to use Vision Transformer as the feature extraction module,uses the RRe Lu activation function to improve the original model,and realizes a new set of image super-resolution generation adversarial network models for small targets: A-ESRGAN.The network is trained on the DIV2 K dataset and tested on the AI-TOD dataset.Through the ablation experiments of various improved modules,compared with the generated data of real-ESRGAN,the image quality of the object area contained in the improved high-resolution dataset has been significantly improved,the peak signal-to-noise ratio of the high-resolution data set generated after the improvement has increased by 2.89 d B.This thesis also uses the FasterRCNN network to detect the image data before and after processing.Compared with the images processed by real-ESRGAN,the images processed by A-ESRGAN contain higher resolution objects,increase the detection AP value by 2%,AP50 The value has increased by 4.8%.Compared with the image of the original unprocessed AI-TOD dataset,the detection AP value has increased by 5.7%,and the AP50 value has increased by 12.3%.Based on the improved A-ESRGAN,this thesis also proposes a repeated reinforcement training algorithm.The algorithm obtains the network parameters of images of different sizes by performing super-resolution reconstruction training of different sizes on the same dataset,and then conducts super-resolution reconstruction training of the same magnification on different datasets,to obtain images with different information amounts,and finally builds a repeated enhancement training algorithm.This method strengthens the utilization of limited data and is suitable for detection scenarios with limited sample data.This thesis compares the improved algorithm with the super-resolution algorithm proposed in recent years on the AI-TOD dataset and Tiny Person dataset,and then uses a number of mainstream target detection network verification algorithms to assist related tasks in the detection scene.The results show that the repeated enhancement training algorithm proposed in this thesis has a better effect.Compared with the original dataset without super-resolution reconstruction processing,the algorithm proposed in this thesis improves the detection AP value of Faster R-CNN by 7.1%,AP50 The value has increased by 17.2%,the detection AP value of SSD-512 has increased by 6.5%,the AP50 value has increased by 17.7%,the detection AP value of YOLOv3 has increased by 6.1%,and the AP50 value has increased by 18.2%.