A ship target detection method based on a multi-scale fusion strategy

Abstract

The invention discloses a ship target detection method based on a multi-scale fusion strategy. The ship target detection method comprises the following steps: S1, preparing a data set for training a feature extraction network model and a target detection network model; S2, designing and obtaining an optimized feature extraction network by combining the concept of multi-scale feature fusion; S3, transplanting the feature extraction network model obtained in the step S2 into a target detection network RFCN to obtain MFF-RFCN based on multi-scale feature fusion; S4, based on the self-built ship target detection data set Dataset-detection in step S1,, carrying out training to obtain a network model which can be better applied to ship target detection. By adopting the technical scheme of the invention, the high-level feature map and the low-level feature map are fused, the semantic information and the position information are integrated, the feature extraction capability of the network canbe further improved in a resource-unlimited use scene, and the method has the advantages of high detection precision, wide coverage scale and the like.

Description

technical field [0001] The invention relates to a ship target detection method at sea, in particular to a ship target detection method based on a multi-scale fusion strategy. Background technique [0002] my country has a vast territorial waters and rich marine minerals. It is a large maritime country, and the ocean is an important barrier to national security. Ship targets are the main body of the ocean, and accurate detection of ship targets is a significant basic work on the road to building a maritime power. [0003] The continuous deepening of research in the field of deep learning has made object classification and detection networks based on deep convolutional neural networks widely used in various fields. However, many deep learning network frameworks generate candidate frames based on a layer of shared feature maps at the end of the feature extraction module and perform target detection and recognition. Although the feature map at the end has strong semantic infor...

AI Technical Summary

Problems solved by technology

[0004] In order to solve this problem, one method is to use data augmentation. During training, the pictures are scaled into different scales and then input into the network for training, but the running time will increase exponentially, and the efficiency of target detection will be greatly reduced. It needs to occupy more video memory capacity, and the requirements for the hardware environment have been improved.

Another method is to solve the memory problem through single-scale training and multi-scale prediction, but this does not fundamentally solve the problem of incomplete semantics and location information in the network feature structure, and the training and prediction goals are not consistent. It is difficult to judge the evaluation index of the training process

Images