Research On Underwater Small Object Detection Method Based On Multi-Scale Feature Fusion

Underwater object detection is one of the key technologies in the visual perception system of autonomous underwater vehicle(AUV).It provides necessary basis for followup object tracking and location.In this paper,the underwater object detection is studied by deep learning method.The underwater object as detection object and multi-scale feature fusion is studied.The underwater detection network model was established and trained,finally the physical experiment verification of the model was completed on the underwater experimental platform.The main work of this paper is as follows:Firstly,this paper introduces the theoretical basis of object detection algorithm based on deep learning.The failure factors of the algorithm to detect some small scale and overlapping targets are summarized and analyzed through experiments.In feature pyramid network,the top-down feature fusion path leads to the imbalance between semantic information and detail information.A bottom-up structure and FCOS feature fusion structure are added to form a path aggregation feature pyramid to enhance the bottom features.The designed multi-scale aggregation feature pyramid is helpful to improve the ability of small object classification and localization.Secondly,in view of the low detection accuracy of small-scale objects when the FCOS algorithm is applied to underwater object detection,the original algorithm is improved.The algorithm first improves the backbone network of FCOS and adds a refined convolution block,which is more conducive to fully extracting the detailed information of the picture.Through multi-scale feature aggregation,features with rich details and strong robustness are generated.The feature map helps to improve the utilization of the underlying features.Then,more recall points are introduced to increase the recall rate of small-scale objects.The GIo U loss function is introduced in the model training to measure the disjointness of the prediction box and the ground-truth box during training.Gradientfree backhaul to speed up model convergence.Finally,in view of the fact that the increased bottom-up aggregation path affects the model detection speed when the MA-FPN algorithm is used for multi-scale feature fusion The feature factor fusion is proposed without considering the fusion ratio of feature layers.Driven by the high-precision detection of underwater objects,the two-stage object detection algorithm Faster R-CNN is optimized.The feature factor fusion mechanism is designed to expand the detailed information of the underlying features.A soft threshold attention mechanism is designed to suppress background information.For objects with similar positions that are mistakenly deleted during non-maximum suppression,the confidence level is used.The weighted fusion prediction box reduces the missed detection of overlapping objects.