Image Target Detection Based On Hierarchical Visual Calculation And Representation Learning

With the rapid development of aerospace and imaging technology,image data acquired through satellite,airborne,and smart devices is increasing.The interpretation of these massive images is infiltrated into various fields such as military and people's livelihood.As one of the basic tasks of interpretation,target detection is of great significance in the fields of military strikes,target tracking,and animal husbandry management.This paper explores and innovates target detection technology from three aspects:(1)Aiming at the aircraft target detection technology of large-scale and complex background optical remote sensing images,this paper proposes a remote monitoring image based on hierarchical vision and prior information for weak surveillance aircraft target detection methods.Optical remote sensing images are characterized by high resolution,large and wide,and the image contains a wealth of information.In this case,posing a huge challenge for the target inspection task.Based on the sparse representation model of the sketch map,this paper proposes a method for extracting potential aircraft sketch line segments based on the angle of the wing by using the positional relationship between the line segment and the line segment in the sketch map.Grayscale information on the surface of the graph,interactively using the sketch map and the original image grayscale image,the line surface combination,further screening the results of the previous step.Constructing parallel line constraints,extracting the symmetric wing.The original grayscale image is again used to achieve aircraft target positioning through region growth.Finally,the experimental comparison with the circumferential frequency filtering aircraft target detection method shows that the weak supervised detection method is superior in aircraft target detection of optical remote sensing images under large and complex background.(2)The paper proposes an algorithm for aircraft candidate box extraction based on sketch map,and combines the Fast R-CNN network to achieve the purpose of detecting aircraft.Firstly,in the candidate box extraction algorithm,the sparse representation model of the sketch map is used to design the potential paired wing sketch line segment extraction method based on the wing leading edge constraint,and different candidate box acquisition methods are designed for aircraft with different angles.The experimental results show that compared with the Selective Search candidate box extraction algorithm,the sketch frame based candidate box extraction method can achieve higher recall rate.Secondly,combined with the Fast R-CNN technology,using the aircraft candidate box obtained in the previous step,the overlapping and multi-scale blocking strategy is designed to train and test the network to realize the aircraft target detection for large optical remote sensing images.(3)This paper proposes a deep weak supervisory target detection and correction method based on the saliency mechanism.In the field of deep learning,the weak supervised target detection method based on the antagonistic complementary learning framework eliminates the need to manually label the object's bounding box,and only uses the image-level label for classification training.The activation map of the network can be used to obtain the target location.However,this type of method still has problems such as low precision of detection results and low overlap between detection result boxes and targets.This paper aims at this problem.A significant mechanism is added to the weakly supervised target detection network based on the activation map.Adding significant loss to the network,guiding the network to learn more features of the target,so that the overlap between the location where the feature map is activated and the location of the target to be detected is improved.Therefore,the accuracy of target detection is improved,and the experimental results compared with other methods demonstrate the effectiveness of the method.