Infrared focal plane detectors play very important roles in the field of military optoelectronic imaging. It has shown unique features in national defense and the national economy, but in the manufacture of infrared focal plane imaging detectors, due to the error in the production process, there are inevitably non-uniform response among the pixels, and some blind pixels that completely lose detecting capabilities will generate the fixed pattern noise in the accessed infrared image. In addition, some flicker pixels whose response values changed violently with time, may affect the imaging quality seriously. How to correct image defects caused by these blind pixels is an important topic in the infrared image processing researches. This paper analyzes the causes of the inhomogeneity of the infrared detector's response and the characteristics of the infrared image, then introduces some popular methods about blind pixels detection and compensation for the infrared detector. The traditional alternative kinds of blind pixels compensation methods often calculate the digital number (DN) value from the neighboring pixel response around the blind pixel position using the single frame image and substitute for the corresponding blind pixels response. However, these methods may lead to a single point of high noise in some specific scenarios. Aiming at such faults, an improved real-time blind pixel compensation method is proposed in the paper. We divide blind pixels into dead pixels and flicker pixels firstly according to the pre-calibrating results using blackbodies. Different algorithms with different thresholds are adopted to detect different types of blind pixels. For the dead pixels, the traditional method is used to replace their responses with their neighboring pixels. For the flicker pixels, a queue consisting of a series of image sequences is built in the memory, temporal filtering is performed for the input image series to reduce the time domain noise. Especially for flicker pixels with contiguous slices, it can better smooth the non-linear error caused by the sharp transition of pixel response with time. For blind pixels on the current frame image, their DN values are replaced by the previous frame sliding filter result in the image sequence queue. In order to simplify the complexity of the hardware design, the temporal upper threshold in the time domain filtering is also set. If the temporal filtering time reaches the upper threshold, the value matching conditions are still not found to compensate for the blind pixels, then the traditional alternative method is used to fill in to ensure the processing time limitation. The algorithm mentioned above and the calculation complexity of hardware implementation are given in detail. Afterwards, the platform for hardware system and the blind pixel correction method described via Verilog HDL and realized on this hardware platform are illustrated. The experimental results show that the algorithm can effectively compensate for the influence of blind pixels on the basis of real-time performance, and plays an important role in the improvement of image quality
This paper proposed a hyperspectral subpixel target detection algorithm based on joint spectral and spatial preprocessing prior to endmember extraction and spectral angle mapping(SAM). Under the condition that the prior information of targets and background is unknown, the spectral and spatial information is used to locate and detect targets. Then we can make hyperspectral subpixel targets detected and recognised. The joint spectral and spatial preprocessing prior to endmember extraction method is performed to extract endmembers. The spectral angle mapping method is used to detect and recognize the interested targets. The hyperspectral image collected by AVIRIS is applied to evaluate the proposed algorithm. The proposed algorithm is compared with SAM algorithm and RX algorithm by a specifically designed experiment. From the results of the experiments, it is illuminated that the proposed algorithm can detect subpixel targets with lower false alarm rate and its performance is better than that of the other algorithms under the same condition.
In order to detect satellite under sky background, we propose an optimized satellite object detection extraction and tracking algorithm under the sky background. The proposed satellite tracking processing consists of two stages. In the first stage of object detection and extraction, the background template based on the mixture Gaussian model is used to establish background frame, and then the background is removed by inter-frame difference method to obtain the object. In the subsequent object tracking stage, this paper proposes an improved untracked Kalman filter algorithm for object tracking. Firstly, it tracks multiple suspected objects in the background, and then introduces a path coherence function to eliminate the false objects. Compared with other methods, the experimental results show that our method can better meet the real-time requirement, eliminate false objects appeared in the sequence of images more efficiently and make the tracking trajectory smoother.
Compared with Support Vector Machine (SVM), Least Squares Support Vector Machine (LS-SVM) has overcome the
shortcoming of higher computational burden by solving linear equations, and has been widely used in classification and
nonlinear function estimation. For dim small targets track predicting in the IR image sequences, a new method based on
LS-SVM is proposed. LS-SVM has prominent advantages in model selecting, over-fitting overcoming and local
minimum overcoming. In this paper, the RBF kernel function is used in LS-SVM, so there are two parameters in
LS-SVM: the regularization parameter γ and the kernel width parameter σ<sup>2</sup>. Since the optimization parameters (γ, σ<sup>2</sup>)
determine the performance of LS-SVM, so their influence on the performance of LS-SVM is analyzed in this paper.
Finally, compared with the Least Square (LS) estimation, the experiments show that LS-SVM can track targets more
precisely and more robustly than LS. Experiments show that the track predicting method based on LS-SVM possesses
the strong learning capability through a small quantity of samples, the good characteristic of generalization and rejection
to random noise. It is a potential track predicting method.