Research Of Ship Detection Based On Remote Sensing Image

Ship detection aims to achieve the localization and recognition of sea-based ships,and has a wide range of applications in fields such as ocean detection,military reconnaissance,and national defense construction.However,due to the difficulties posed by ships such as the diversity of orientation angles and significant scale changes,traditional horizontal bounding boxes result in identifying redundant regions and limits the accuracy of depiction and characterization of ship targets.Consequently,the application of general object detection methods is restricted in ship detection tasks.To address the above issues,the researchers proposed using bounding boxes with rotation angles to locate ships.While this approach has enabled ship detection in some cases,significant challenges remain: 1)The IoU loss of the rotating bounding box used in ship detection tasks is non-differentiable,creating difficulties in applying gradient backpropagation to the ship detection network.2)Existing convolutional neural networks have a limited field of view,which leads to difficulties in accurately aligning ship features distributed in multiple directions.3)When accounting for rotation angles,the number of preset anchor boxes in the detection network is significantly greater than for traditional horizontal boxes,resulting in extremely high computational costs.This thesis proposes a series of studies aimed to mitigate the aforementioned problems.(1)Regarding loss design,a solution is proposed to address the non-differentiable issue of the IoU of rotating bounding boxes,using a neural network to fit the IoU loss and establish an IoU fitting loss for rotating ships.The proposed method constructs an IoU fitting module based on a multi-layer perceptron,and designs a random rotating bounding box generation algorithm and a rotating bounding box normalization algorithm,effectively achieving precise fitting of IoU losses.The proposed method outperforms the baseline method by achieving a 1.91% improvement on mAP.(2)In terms of feature alignment,a solution is proposed for the problem of horizontal fields of view failing to adequately represent ships distributed in multiple directions.A dynamic direction alignment network is established for ships with multi-directional distributions through dynamic field-of-view adjustments using rotating anchor boxes.The proposed method designs a dynamic direction alignment convolution module,as well as constructing fine-grained feature modules and dynamic target prediction heads through dynamic direction alignment,which enables the comprehensive and precise exploitation of the characteristics of rotating ships.Compared to the baseline network,the proposed approach yields a 6.21% increase on mAP.(3)For model acceleration,in response to the expensive computational cost caused by redundant rotation anchor boxes,a solution using sparse proposal boxes instead of dense anchor boxes is proposed.An efficient sparse prediction model is established for ship detection.Specifically,The proposed method constructs a two-stage rotation feature decoder based on progressive learning strategy.To optimize model performance,the method employs a matching loss and label assignment strategy that leverages IoU fitting loss to facilitate accurate object detection.Moreover,the model utilizes a dynamic feature interaction prediction head constructed via dynamic direction alignment convolution to enhance detection accuracy,and realizes end-to-end training and inference of the model.Compared with existing methods,the proposed method reduces FLOPs by 32.45% while simultaneously improving the mAP by 2.48%.