Multimedia Tools and Applications
the target regions and centroid positions as the reference of evaluation. We test the
network on real images trained only by the synthetic dataset and it proves that the
synthetic images have similar features as real star images.
– We design and train a lightweight fully convolutional neural network to get the pixel-wise
segmentation result of a single frame. This approach make use of the features of the
image in different scales to achieve end-to-end target-background separation rapidly.
– The proposed network model is validated on different star image datasets experimen-
tally. Compared with the state-of-the-art methods, our proposed model achieves a better
background suppressing and target detection performance, especially higher detection
rate and less false alarms caused by noises.
The remaining of this paper is organized as follows. Firstly, the related works on single-
frame small target detection are presented in Section 2. Secondly, our proposed method,
including the architecture, training of the network and the dataset generation are stated in
Section 3. Finally, the experimental results and the conclusions are elaborated in Section 4.
2 Related work
Single-frame small target detection The previous proposed single-frame small space tar-
get detection methods could be roughly classified into two categories, multistage target
detection including a series of pre-processing and single-stage target detection that separate
the target from image directly.
In the first type, in order to remove the distraction of the noises and increase the SNR of
the dim targets, variety of denoising methods are designed based on the features of different
kinds of noises. These methods perform step by step to remove the interference of each
kinds of noises. The noise spikes, such as hot-pixels, are identified and removed by their
common feature that single pixel with exceptional brightness far above the main distribution
of background noises [9, 16]. The cosmic rays are removed by their sharp edges [24, 34]and
the random location where they appears in different images [35]. The uneven background
illumination are fitted and subtracted from the original image [16] based on the fact that
the image background is very smooth and can be modeled with polynomial fits. Benefited
to the denoising process, the targets are enhanced, and then can be detected much easier by
the matched filtering techniques in different scales [5, 16] or other threshold based methods
[8, 18, 27]. Although the multistage process achieve optimal results on each step, the wrong
detection of noises may lead to artifacts in the following steps, especially when the noise
spikes are close to the targets. The processing speed of single frame also slow down when
all these noises appear together.
The second type of single-frame small target detection methods focus on the single
stage target-background separation, which has not been explored extensively. Some meth-
ods exploit the non-local self-correlation property of background patches, assuming that all
background patches come from a single subspace or a mixture of low-rank subspace clus-
ters, so target-background separation can be realized with the low rank matrix recovery [4,
40]. To correctly detect the small target located in a highly heterogeneous background, He
et al. [13] proposed a low-rank and sparse representation model under the multi-subspace-
cluster assumption. Essentially, this type of methods attempts to model the small targets
as outliers in the input data, but the spike noises cannot be distinguished from the target
in many cases. Moreover, inspired by the semantic segmentation method based on extract-
ing distinguishable features, we have applied texton-boost [28] on small target detection