Since primitive line segments extraction is an important procedure

how to select the proper line segments extraction approach is the key to the extraction of linear object of SAR images.This paper first briefly reviews three existing methods: Radon technique

the Burns line detector

the line detector based on template matching.In the Radon Transform methods the transform is applied to each sets of the connected pixels

the relative maximum in the transform corresponds to possible lines in the binary image.In order to reduce inter-edge influence

after extracting the possible primitive line segment

only the line associated with the maximum Radon coefficient is inspected at each iteration instead of continuing and proceeding with the next relative maximum.This process is recursively repeated until the maximum Radon coefficient reaches a minimum value.In Burns line detection method

edge pixels are first determined by convolution with two simple 2×2 masks.The pixels are grouped into line-support regions of similar gradient orientation.The intensity surface associated with each line-support region is approximated by a planar surface.Straight lines are extracted by intersecting this fitted plane with a horizontal plane representing the average intensity of the region weighted by a local gradient magnitude.In the method based on template matching

each edge pixel is given a line label based on certain templates in the first step.The result of this step is that edge pixels that lie on the same line get the same line label.Furthermore

corresponding to each line label

a complete description of the line is obtained.In the second step

these lines are further linked based on certain criteria.While an extensive literature exists on the problem of PL extraction

most of the researches only give qualitative analysis of the behavior of PL extraction

made by a visual comparison of different results.To provide a solution

in this paper we propose a set of quantitative evaluation metrics to determine how many lines are correctly extracted.The basic idea is to compare the extracted PLs with the lines in the ground truth data

thus the performance of the PL extraction can be showed through a series of performance metrics.We consider in this paper three categories of performance metrics

namely

a) pixel-based metrics

which are composed of completeness

correctness

quality

and ROC curve;b) line-based metrics

which are composed of the percentage of the extracted lines and mean number of PLs per line

c) computational complexity.To correctly compute these quantities

we first need to define when a line may be considered as extracted

and then manually extract the start and end points of the center line of each line in the test image as a reference.In summary

the paper studies approaches of primitive line segments extraction

presents a series of quantitative performance evaluation measures and evaluates these three methods.The proper line segments extraction approach for linear object extraction can be selected based on the evaluation results.