Research On Infrared UAV Target Detection Technology Based On Machine Learnin

Target detection technology is an important branch of computer vision,and in recent years,with the rapid development of machine learning technology,convolutional neural network has become the core method to solve the current target detection problem.In the field of target detection,the detection of infrared UAV targets has been the focus and difficulty.Due to the small number of target pixels and weak features,it is difficult for traditional detection networks to learn effective target features.Image super-resolution technology can increase the resolution of the target and enrich the target features,and the combination of deep neural network and image super-resolution technology can effectively improve the detection accuracy of small infrared targets.In this paper,we produce an infrared UAV target dataset with images containing many different backgrounds and UAV targets covering multiple scales.The infrared UAV data enrichment algorithm and target detection algorithm are studied in depth based on this dataset,and the main work and innovation points of the paper are as follows:(1)A data augmentation algorithm based on an improved deep convolutional generative adversarial network(DCGAN)is proposed.Firstly,the UAV targets are extracted to train the DCGAN,and new targets different from the training set samples can be obtained by the improved generator.After the background is removed by the new target extraction module,the new target is expanded to the original dataset,so that the infrared image which originally has only a single target contains multiple targets,and the target augmentation of the dataset is achieved.The improved DCGAN generator incorporates a new residual module to generate clearer UAV images,and the new target extraction module combines the image segmentation algorithm to remove the background of the generated images before target augmentation,making the augmented dataset images more accurate.After comparing the experiments with the detection algorithm,the augmented dataset improves the average precision(AP)of SSD algorithm by 0.014 and Faster-Rcnn algorithm by 0.013,which fully proves the effectiveness of the proposed data augmentation algorithm.(2)A new algorithm for infrared small target detection is proposed: the Regional Super Resolution Generative Adversarial Network(RSRGAN).The algorithm consists of a lightweight network with a simple structure and a regional context network(RCN)as the backbone network,which consumes less computational cost to extract the suspected regions that may contain targets.The generator of the super-resolution module consists of two parts:distribution conversion and super-resolution enhancement,which first converts a small blurred IR UAV target into a clear target and then performs resolution enhancement to obtain a better enhancement effect.The enhanced target has richer detail features,and then the target can be accurately identified and localized through the secondary detection of the fully connected layer.The front-end RCN backbone network and the new distribution transformation generator make the super-resolution network better combined in the target detection algorithm framework,and the coarse extraction plus fine detection method makes the algorithm have higher accuracy rate.After experimental verification,the algorithm of this paper achieves higher average accuracy than the commonly used target detection algorithms on homemade IR UAV small target data,and improves the AP value by 0.016 compared with Efficient Det algorithm,which fully proves that the proposed IR UAV small target detection algorithm has more accurate and stable detection effect.