Optimization method of Tiny-YOLO network for detecting ship target on satellite

Abstract

The invention discloses an optimization method of a Tiny-YOLO network for detecting a ship target on a satellite. The method comprises the following steps: employing a sample set of a ship image, carrying out the training of an original Tiny-YOLO network, and obtaining the parameters of a convolution kernel in each convolution layer in the network; determining a Tiny-YOLO network for ship target detection according to the original Tiny-YOLO network and the parameters of the convolution kernel in each convolution layer in the network; according to the method, the Tiny-YOLO network is sparsifiedby reducing convolution kernels, and transfer learning is carried out according to the position of each convolution layer of the sparsified Tiny-YOLO network, so that the operation speed and the detection accuracy of the sparsified Tiny-YOLO network on a satellite meet the requirements; and the convolution kernel parameters in the Tiny-YOLO network after transfer learning are converted into integers from floating-point numbers, so that a final Tiny-YOLO network can be obtained, and therefore, the requirement of improving an operation speed by using the improved Tiny-YOLO network on a satellite can be satisfied.

Description

technical field [0001] The invention relates to an optimization method of a Tiny-YOLO network used for ship target detection on a star, and belongs to the technical field of convolutional neural network structure optimization. Background technique [0002] Most of the current ship target detection algorithms are based on Synthetic Aperture Radar (SAR) or infrared images, such as literature [1]. In comparison, the research on detection algorithms based on satellite optical remote sensing images started relatively late. For target detection, the traditional method is based on manual design and extraction of target features. At present, it has developed into a more advanced method based on deep learning, such as Faster RCNN. This method does not need to artificially design the characteristics of the target, design the network for detection, complete the training of the detection network through a large number of samples, and obtain good detection results when the target is in a...

AI Technical Summary

Problems solved by technology

This method provides an idea for the realization of on-board applications, but the network processed by this party can only achieve target classification and cannot achieve target detection

Images