Sub-pixel target detection is one of the key technologies in the applications of hyperspectral images.Since the high dimensions of hyperspectral data increase apparently the storage space and complexity of data processing,real-time processing has become a crucial problem for target detection.Adaptive matched filter (AMF) is an effective algorithm for sub-pixel target detection.This paper derived the real-time AMF target detection procedure of hyperspectral images by using AMF as the sub-pixel target detection algorithm,based on the realization of real-time inversing of hyperspectral data’s covariance matrix with the pixel-by-pixel format transmission and storage by using Woodbury lemma.Expe-riments were conducted on synthetic data and real hyperspectral images.The results demonstrate that compared with non-real time AMF,real-time AMF needs less storage space and can achieve the same or slightly better detection accuracy.

Covariate shift incurred by occlusion and illumination variations is an important problem for real-world face recognition systems.This paper explored this problem from the perspective of dictionary coding.By reviewing several extant structured error coding methods,this paper indicated that these error coding methods can be rewritten as a linear system by combining training dictionary and well-designed occlusion dictionary.Due to the importance of occlusion dictionary in structured error coding,this paper studied the dictionary learning method,K-SVD (Singular Value Decomposition),which is used in the Gabor feature based robust representation and classification (GRRC) method,and has been paid great attentions in the field of error coding.The K-SVD learned occlusion dictionary is strongly redundant and lack of natural structures.In addition,K-SVD is time-consuming.This paper proposed an SVD-based occlusion dictionary learning method.It is simple,but generates a more compacted and structured occlusion dictionary.Experiments on three face datasets,including Extended Yale B,UMBDB and AR,demonstrates that the proposed SVD-based GRRC consis-tently outperforms the K-SVD-based GRRC in several challenging situations.

Traditional maximum margin criterion usually ignores the differences between classes in the computation of the between-class scatter matrix.However,for facial age estimation,the differences between age labels are very significant.Therefore,this paper proposed a novel dimensionality reduction algorithm,called label-sensitive maximum margin criterion (lsMMC),by introducing a distance metric between the classes.In addition,considering the complicated facial aging process,this paper proposed a two-steps local regression algorithm named K nearest neighbors-label distribution support vector regressor (KNN-LDSVR) for age estimation.The mean absolute error of the proposed facial aging estimation method on the FGNET database subset is 4.1 years,which improves the performance compared with existing age estimation methods.

Aiming at the problem of the moving cast shadow in the video surveillance,this paper proposed an shadow removal algorithm which combines color feature,normalized vector distance and intensity ratio.First,the background picture is built according to Gaussian mixture model,and motion region is acquired by background subtraction.Then,serial fusion method is adapted to detect and remove shadow pixels.Based on shadow detection according to the color consistent feature in RGB color space,the normalized vector distance distribution histogram is implemented to detect sha-dow pixels further.Finally,in view of the mistaken identification in the testing process,the illumination model of pixel is built and the intensity ratio of shadow pixel and background pixel is calculated to rule out the mistakenly identified foreground pixels according to the confidence interval.The results of experiment show that the proposed method can overcome the limitation of single feature method,and is able to detect and remove shadow under various circumstances efficiently.The adaptability and robustness of this algorithm are validated,and its processing time is moderate.

Due to the complexity of the urine sediment space environment,there is much redundant information of the collected tangible component image,and it also becomes difficult to extract effective image information.Therefore,the amount of data that need to be deal with is huge.Although the serial version DJ8000 system platform of BP neural network algorithm solves the problem of recognition accuracy of tangible components such as cells,it can’t meet the real-time requirement of urine sediment image medical examination.To solve this problem,this paper presented a system platform of parallel processing GPU framework based on BP neural network algorithm optimization.It uses parallel optimization framework to synchronize and accelerate processing of data efficiently.At the same time,it supports the hardware platform based on GPU computing and test platform.Whether from the hardware indicators,data transmission and bus technology or hardware and software compatibility,it will help solve the problems,which often occur in the uneven load irregularities.Experimental data show that BP neural network algorithm for urinary sediment identification improve the performance parameters such as speedup,aging ratio and running time on GPU platform processing platform.Compared with DJ8000 system platform,the parallel processing GPU framework system platforms of AMD HD7970 and NVIDAGTX680 are optimized,and their corresponding acceleration ratio parameter values are 10.82~21.35 and 7.63~15.28 standard equivalents respectively.The data show that optimizing the mapping relationship between logical data,address data and linear seeking function in the GPU processing system of parallel frame can dynamically allocate and optimize the algorithm structure and optimize the mapping between software and hardware system.Finally,it solves the problem of performance bottlenecks caused by load imbalance between threads.Thus,it effectively resolves the problem of real-time detection in urinary sediment environment.

The off-line training and the high dimension of facial features in face recognition lead to the difficulty of achieving real-time processing performance.To solve this problem,the local fusion features and the hierarchical incremental tree were applied to construct a fast face recognition algorithm.Firstly,the supervised descent method(SDM) is used to locate the facial feature points.The feature of multi block-center symmetric local binary patterns(MB-CSLBP) in the neighborhood of each facial feature point is extracted and fused in series,which constitutes the proposed facial feature of local fusion feature of MB-CSLBP(LFP-MB-CSLBP).Then the above facial feature is sent into hierarchical incremental tree(HI-tree).Because the hierarchical clustering algorithm is used in the HI-tree to achieve incremental learning,it can train the recognition model online.Finally,the recognition rate and consuming time of the proposed algorithm are tested on three face databases and real application of video-based face recognition.The experimental results show that the proposed algorithm has better real-time computation and accuracy compared with other current approaches.

Detection and recognition of traffic light are important for driverless cars and advanced driver assistance systems(ADAS).In order to satisfy the requirements of traffic light detection and recognition in complex urban environment,a real-time detection and recognition algorithm based on time-space model (TSM) was proposed.It was established based on thetime-space continuous variation relationship of video-frame sequence.The proposed algorithm consists of three parts.The first part is fast image segmentation and compression algorithm based on color,which is used to improve the computational efficiency.Second,time-space model of multi-frame image sequence is introduced to improve the accuracy of detection stage.Third,recognition of traffic lights is achieved by using support vector machine (SVM) with histogram of oriented gradients (HOG) features.Experiment results show that this novel algorithm has strong robustness,high efficiency and accuracy.