Research On Aerial Image Object Detection Method Based On Full Convolution Neural Networks

Research on high-resolution image object detection algorithms based on large-scale data is an important direction.At present,the algorithm is roughly divided into two types:traditional algorithms and deep learning algorithms.Traditional algorithms are difficult to apply in large-scale data.It extract feature through manual operators.Its feature expression ability is limited and redundancy.Manual operators rely on professional domain knowledge,and it is difficult to train an effective feature extractor from large-scale data,and the relationship between data cannot be fully exploited.The deep learning algorithm has efficient feature extraction ability,powerful feature expression ability and learning optimization ability,which can provide support for image feature extraction algorithm research.Therefore,for aerial image detection problems,deep learning algorithms have advantages that are incomparable with traditional algorithms in terms of speed and accuracy.In this paper,based on the full convolutional neural network object detection algorithm,the related research on aviation detection scenarios will be carried out.The main research contents are as follows.In this paper,a heuristic training strategy is proposed for the scale imbalance problem in aviation detection scenarios.This strategy generates new training samples according to the prior information of the object,changes the overall distribution of object scales,and alleviates the scale imbalance problem.This paper proposes an anchor clustering algorithm,which considers the degree of overlap between the sample and the class center and the degree of consistency of the length and width ratios.Compared with the uniformly set anchor scale,the algorithm can obtain more accurate anchor scale according to the result.In this paper,a global optimization method is proposed for the training supervision method of the detector.The global distribution consistency error is used to optimize the network parameters.The global distribution consistency means that the distribution of the training samples is consistent with the distribution of the model prediction results,and the method is more general.The heuristic training strategy effectively solve the scale imbalance problem,and the performance is improved by 3.4 percentages compared with the baseline.The an-chor clustering method makes the detector's prediction more accurate,and the detector improves 1 percentage.The global optimization method not only improves the algorithm on the aviation scene dataset,but also validates its generalization ability.This paper also conducts comparative experiments in the natural scene dataset.This method makes some classical algorithms more than 1 percentage.The detection algorithm performance has obtained 74.2 percentages,which is very competitive with other algorithms in the academia,especially in the small-scale object.