Font Size: a A A

Research And Application Of Multi-dimensional Sparse Signal Recovery Algorithm

Posted on:2016-06-12Degree:DoctorType:Dissertation
Country:ChinaCandidate:Y Q HuoFull Text:PDF
GTID:1228330461466846Subject:Agricultural Electrification and Automation
Abstract/Summary:PDF Full Text Request
Information Technology(IT) is the key support technology to implement the precision agriculture. With the continuous development in modern agricultural technology, massive multi-dimensional data such as image, video, and hyperspectral data are becoming increasingly ubiquitous across science and engineering, which increase in the computational complexity and memory usage and bring the challage to the encoder and decoder. How to take full advantage of the intrinsic feature of the multi-dimensional signal to reduce the computational complexity and memory usage, it is the basic premise to the precision agriculture.To solve the low recovery quality and high complexity problem, this paper proposes the Two Dimensional Subspace Pursuit(2DSP) algorithm and Three Dimensional Orthogonal Matching Pursuit(3D-OMP) algorithm to improve the performance of the multi-dimensional signal recovery algorithm and accelerates the 3D-OMP algorithm based on the many-core processor Graphics Processing Unit(GPU) to enhance the running speed of the decoding algorithm even more. In addition, Compressive Sensing(CS) theory has been succeeded in apple fruit disease pattern recognition in agricultural engineering. The author’s main jobs and contributions are outlined as follows:(1) To solve the low quality problem of the 2D-OMP algorithm for the recovery of 2D sparse signals, the 2DSP algorithm is proposed which pursuits the subspace spanned by 2D atoms instead of the most matched 2D atom. Hence it improves the performance of the decoder while the time complexity and space complexity is indentical to the 2D-OMP. The experimental results reveal that 2DSP recovery algorithm is able to exactly reconstruct the synthetic 2D sparse signal. Further more, 2DSP is superior to 2D-OMP algorithm when reconstructs the real 2D image, and its PSNR is about 0.5d B higher than that of 2D-OMP.(2) Three Dimensional Separable Operator(3DSO) method is proposed which solves the high computational complexity problem for the 3D signal measurement at encoder where the 3D signal processed by the measurement matrix and sparsifying base on the x, y and z dimension separately. The detailed mathematical deduction shows that the proposed method significantly reduces the size of the measurement marix and sparsifying base compared with the traditional global measurement, and its time complexity and space complexity are both3O(mn) roughly 21 m of that of global measurement while maintains the identical effect to the golobal measurement. More over, the mathematical analysis of the mutual coherence between the measurement matrix and spasifying base and the RIP property are provided in detail.(3) The 3D-OMP recovery algorithm is proposed to solve the high complexity problem for the reconstruction of three dimensional sparse signal which takes full advantage of the high redundant characteristic of a 3D signal. In the propose algorithm the 3D signal is represented as a weighted sum of 3D atoms where a best matched 3D atom is searched in each iterator to be added to the 3D support set, and then the 3D sparse signal is reconstructed via the least square method based on the k matched 3D atoms in the support set. The theoretical analysis shows that the time complexity and space complexity are 3O(mn) and3O(n) respectively. The experimental results indicates that 3D-OMP algorithm can perfectly reconstruct the 3D synthetic sparse signal with high probability and outperforms that of KCS algorithm in the recovery quality and probability of perfect reconstruction while achieving lower performance than that of the global measurement. Further more, the 3D-OMP algorithm has lower complexity and faster running speed than that of KCS when adopting the real hyperspectral image to verify algorithms.(4) The 3D-OMP algorithm is still time-consuming on the account of the high density of a 3D data. However, the theoretical analysis reveals that the 3D-OMP algorithm has high parallelism degree due to the amount of the matrix manipulation. Accordingly the parallel 3D-OMP algorithm is proposed based on GPU utilizing the powerful parallel data-processing ability to accelerate the recovery algorithm. Further more, some optimizing strategies are adopted, e.g., memory accessing optimization, reduction, instruction-level optimization, task dividing, etc. Experimental results show that the parallel 3D projection which is the most time-comsuming part of the 3D-OMP algorithm can achieve 390 times over its original C code optimized by the 2o, and the parallel residual part gains 55 times. In consequence, the parallel 3D-OMP algorithm can achieve about 146 times speedup on GPU compared with the sequential C code on CPU, which releases the computational burden significantly at decoder.(5)The disease pattern recognition model base on Compressive Sensing theory is proposed for apple fruit disease classification in the precision agriculture. The main idea of the model is that one disease sample can be represented as the weighted sum of the other samples of the same sort and the weighted coefficient is sparse. According to this theory, the disease pattern recognition model is built based on CS, and then coefficients of the test sample projected on the feature matrix are solved via the standard CS recovery algorithm. Further more, the classification of samples is completed according to coefficients. The experiment involved three common apple fruit diseases, e.g., ring spot, anthracnose, and new ring spot, etc, where there are 78 images in total to be choosed and each cartegory has 26 images. Mean while, the SVM recogonition model is realized for the comparison with the proposed mode. Finally, experimental results show that the correct recognition rate of each disease based on the CS recognition model is 80%, 90%, and 100% respectively, and the average correct recognition is 90%, which is equal to that of the SVM model where the correct recognition rate is 80%, 100%, and 90% respectively, and the average correct recognition is 90%. Hence it can be drawn that the proposed recognition model based on CS can effectively recognize the apple fruit disease and has a certain practical value in agricultural disease recognition.
Keywords/Search Tags:Two dimensional subspace pursuit, three dimensional separable operator, three dimensional orthogonal mathing pursuit, many core processor, disease recognition
PDF Full Text Request
Related items