| Target detection technology has always been a popular research direction in the field of computer vision.Target detection technology has been widely used in both civil and military fields.Mobile Agent,the entity having a certain hardware computing capability.The mobile agents in the civil field mainly include carriers such as mobile robots,unmanned vehicles,and drones.Target detection application scenarios for civil agents include object detection and pedestrian detection in natural scenes.The mobile agents in the military field mainly include satellites,missiles,etc.The target detection application scenarios of the military agents include target detection in satellite-oriented remote sensing images and target detection of airborne equipment.With the rise of 5G and the Internet of Things technology,the expansion of application scenarios and changes in application conditions have placed higher requirements on target detection technology.However,no matter which kind of algorithm is used,traditional target detection algorithms are inseparable from a large number of training samples.When the application scenario of the algorithm does not allow obtaining more samples or the application scenario requires high performance of the algorithm.These algorithms are not suitable for directly dealing with small sample problems,and their limitations are mainly reflected in the dual decline in detection recall and accuracy.In the case of a small number of existing samples,how to improve the accuracy of small sample target detection as much as possible and reduce false alarms in the results are the main issues discussed in this article.In view of the above problems,this paper proposes a small sample target detection solution for mobile agents.This solution includes research on target sample augmentation technology based on existing samples and implementation of lightweight small sample target detection algorithms.The main work and innovations are as follows:(1)The traditional non-deep learning-based feature extraction algorithm has the advantages of high speed and small size,but it is extremely vulnerable to background information interference during the detection process,resulting in a high error rate.Feature extraction algorithms based on large-scale deep learning have higher feature extraction capabilities,but they have higher hardware requirements.For mobile agents with limited computing capabilities,traditional deep learning-based object detection algorithms have obvious limitations.Aiming at this problem,this paper proposes a small sample target detection algorithm for mobile agents.This algorithm uses feature extractors based on traditional methods to extract features such as texture and gradient of the target,and then uses a shallow convolutional neural network to construct a secondary filtering network,which is used to remove erroneous extraction results.Experiments show that the algorithm has good performance in small sample detection while maintaining high robustness.(2)This paper explores a small sample-oriented target sample augmentation method,which uses a traditional sample augmentation method and a deep convolutional adversarial neural network-based target sample generation method.Aiming at the problems of traditional deep convolutional adversarial neural networks such as poor robustness,difficult training,and easy disappearance of gradients,the idea of variational discriminator bottleneck(VDB)was introduced,and the discriminator’s discriminating ability was restricted during the training process to achieve good generation effect.In the experimental stage,the effectiveness of the target samples generated by the sample augmentation method used in this paper is evaluated,which to find the optimal sample combination and training method.(3)According to the research results of this paper,a small sample target detection platform based on C ++ language is constructed.The platform requires low computing power of hardware.The platform is small,fast,and easy to deploy,which is means a lot to mobile agent.The final test results show that the algorithm in this paper has a good applicability in the application scenario of simply using the CPU. |
