Research On Key Techniques Of Object Detection And Recognition Based On Convolutional Neural Networks And Embedded Systems

Appliying object detection and recognition algorithms based on convolutional neural network on an embedded platform is an important technical means to realize intelligent photoelectric imaging detection.This article focuses on the current gap between the theoretical research of convolutional neural network and the actual needs of photoelectric imaging detection.From this realistic background,the key theoretical and technical problems in both model training and model deployment stages under the embedded system during the actual task of photoelectric imaging detection using convolutional neural network models are analyzed,including the contradiction between training reliance on large-scale samples and the lack of sample data in actual application scenarios,the contradiction between the real-time requirements of engineering applications and the insufficient computing power of embedded platforms,and the contradictions between the classic object detection frameworks and diverse test object requirements.Focusing on these problems and contradictions,researches have been carried out on small sample dataset augmentation methods and convolutional neural network training strategies,model pruning compression methods,and improvement of proposing methods for regions of interest.1.A domain adaptation method based on generative adversarial networks is proposed to solve the contradiction between the dependence of convolutional neural network model training on large-scale samples and the lack of sample data in practical application scenarios.For the case where the training error of the convolutional neural network in the source domain is low but the generalization error in the target domain is high,this method first expands the target domain data set by generating an unlabeled data sample by generating an adversarial network,and then uses a weighted distributed pseudo-label training strategy,using the augmented target domain dataset to transfer training the convolutional neural network model.Experimental results show that this method can improve the generalization ability of the convolutional neural network model in the target domain with a small number of samples in the target domain and no labeled information.2.A domain adaptive method based on adversarial autoencoder is proposed.This method addresses the problem that distributed pseudo-labels are unreliable when the quality of unlabeled samples generated by the adversarial network is too low or the generalization ability of the initial convolutional neural network model is too low.The data is augmented by directly generating labeled target domain samples.method.By combining the autoencoder and the generative adversarial network,and introducing a generative model training strategy based on partial parameter sharing,the labeled source domain image is converted into the target domain style,and the target domain data obtained through the style conversion is used for transfer training the convolutional neural network models.Experiments show that this method can effectively improve the generalization ability of the convolutional neural network model.3.An object detection method based on arbitrary oriatnted regions of interest is proposed.This method aims to solve the problem that the detection method based on the classical axis aligned region of interest does not adapt to the characteristics of the object to be tested in a long-range overlooking application scenario.This method defines a region of interest bounding box with an arbitrary angle,and uses an arbitrary orieanted region of interest feature mapping method based on sub-regional Ro I align to avoid the situation of a large amount of background interference information in the region of interest bounding box.Experiments prove that this method is particularly suitable for high-density and small targets detection in overhead scenarios such as drones and remote sensing.4.Sparse training-based pruning and quantization method is being studied to reduce the amount of calculation during model inference,aiming to solve the contradiction between the real-time requirements of engineering applications and the insufficient computing power of embedded platforms.Through model pruning and quantization,the reasoning speed of the detection model is greatly improved while ensuring the accuracy of the model as much as possible.In two different photoelectric imaging detection applications,the methods proposed in this thesis have been used to complete the training of object detection and recognition models based on convolutional neural networks,and the models have been deployed on embedded platforms of different architectures.Engineering applications have proved the effectiveness of the proposed methods.