Survey of Ship Detection in SAR Images Based on Deep Learning

Xiaohan Hou; Guodong Jin; Lining Tan

doi:10.3788/LOP202158.0400005

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 4 >Page 0400005 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 4, 0400005 (2021)

Survey of Ship Detection in SAR Images Based on Deep Learning

Xiaohan Hou^*, Guodong Jin^*, and Lining Tan

Author Affiliations

College of Nuclear Engineering, Rocket Army Engineering University, Xi’an, Shaanxi 710025, China

show less

DOI: 10.3788/LOP202158.0400005 Cite this Article Set citation alerts

Xiaohan Hou, Guodong Jin, Lining Tan. Survey of Ship Detection in SAR Images Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2021, 58(4): 0400005 Copy Citation Text

show less

Fig. 1. Flowchart of two-stage target detection algorithm

Download full size

Fig. 2. Flowchart of single stage target detection algorithm

Download full size

Fig. 3. Processing method of dataset expansion

Download full size

Fig. 4. GaN processing process

Download full size

Fig. 5. Framework of FPN

Download full size

Fig. 6. Structure of hyperdense connection network

Download full size

Fig. 7. Framework of CBAM

Download full size

Fig. 8. Framework for context fusion algorithm

Download full size

Fig. 9. Different target borders. (a) Target minimum outbound vertical border; (b) target minimum outbound rectangular border

Download full size

Dataset	Open SAR ship	SSDD	SAR-ship-dataset	AIR-SARship-1.0
Data information	There are 11346 slices, including ten types of ships	There are 1160 slices, including 2456 ships	There are 43819 slices, including 59535 ships	There are 31 scenes in Gaofen No. 3 SAR image
Advantage	1^st ship target dataset to provide benchmark data for researchers in this field	The dataset contains SAR images with different resolution, polarization, sea condition, large sea area and landing conditions	There has been a large-scale increase in the amount of data and the types of ships	Including ports, islands and reefs, different levels of sea conditions, the background covers variety of scenes, such as inshore and offshore
Shortcoming	The number of samples among categories is not balanced, so it is difficult to train a better classification model	The amount of data is still lacking	Most of the datasets are offshore background, the nearshore background is less, and the background is relatively simple	The types of vessels are still civilian ships

Table 1. Existing SAR ship datasets

Method		Operation method	Advantage	Shortcoming
Geometric transformation	Translation	G'₍_i'_,_j'₎=G₍_i₊_x_,_j₊_y₎	Enrich the position and scale of objects in the image	The effect is not ideal in location-sensitive tasks
	Rotation	W'=Wcos θ+Qsin θQ'=-Wsin θ+Qcos θ
	Cutting	Random image clipping using a certain overlap ratio
	Cutting	Random image clipping using a certain overlap ratio		After cropping, the output needs to be scaled to a fixed size, which may cause image distortion
	Zoom	Randomly select the expansion scale, place the original image in the lower right corner of the expansion image, and fill other blank areas with channel mean
Optical transformation	Hue	Randomly add a real number to each point in the image with a probability of 0.5	Add images under different lights and scenes	In the case that the edge of the SAR image itself is not clear, it may cause the model to be difficult to converge
Optical transformation	Saturation degree	Randomly multiply each point in the image by a real number in the HSV gamut space	Add images under different lights and scenes
Increase noise	Gaussian/salt-and- pepper noise	G'₍_i'_,_j'₎=G₍_i_,_j₎+NG'₍_i'_,_j'₎=G₍_i_,_j₎·N	Generate robustness to natural disturbances and improve the generalization ability of the model	Excessive noise affects the output of the model
Data source expansion	Change the background	Combine detected objects with other background images	Increase the richness of the dataset	The morphological characteristics of the target itself have not been changed

Table 2. Comparison of traditional data augmentation methods

Category	Method	Advantage	Shortcoming
Target feature extraction	Characteristic pyramid	The problem of multi-scale detection is improved	The lower sampling rate is large, which makes the edge of the object difficult to predict, increases the difficulty of returning to the boundary, and multiple up-sampling operations increase the difficulty of detection
	Super dense connection	The problem of gradient disappearance has been solved to some extent;a large number of features are reused and the number of channels of the feature graph is small	Multiple data replication is required, and certain video memory optimization techniques are needed
	CBAM	Improve the accuracy of detection	The speed has an impact
	Context fusion	Integrate background information, model the relationship between objects, and improve the understanding of the scene	Introduce non-target information
Target border design	Direction design	Fit the multi-angle characteristics of warship	Additional parameters need to be redesigned
Target border design	Scale design	The anchor frame is more in line with the characteristics of the dataset itself	The more the number of anchor involved in the clustering process, the higher the accuracy, but at the same time, it will increase the amount of computation

Table 3. Comparison of algorithms in SAR ship target detection

Xiaohan Hou, Guodong Jin, Lining Tan. Survey of Ship Detection in SAR Images Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2021, 58(4): 0400005

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information