| With the development of spectral detection technology and photoelectric imaging technology,infrared spectroscopy is widely used in many industries.However,in the process of acquiring infrared spectroscopy signals,infrared spectrum often suffers from much interference,such as the aging of infrared spectrometers,naturally random noise and other factors,thus,the infrared spectrum is degraded.So,the deconvolution method is proposed to solve this problem,but the traditional deconvolution method has its own limitations more or less.These limitations are mainly reflected in: a single processing mode,relying on a priori knowledge,manual design parameters,and the speed of terative optimization is slow,so it is urgent to propose a reasonable deconvolution method.This thesis focuses on the research of the deconvolution method of infrared spectrum signals under the degradation of a single fuzzy kernel and multiple fuzzy kernels.The main work is as follows:(1)Based on the theory of deep learning,a dual-stream neural network is proposed to reconstruct clean infrared spectrum from degraded infrared spectrum directly.This network structure enhances the ability of spectrum to extract features in peaks,valleys and flat areas.Meanwhile,in order to reduce the parameter redundancy and the difficulty of training in solving the gradient,a new activation function is proposed to merge this dual-flow neural network structure into an activation function in the form of a function.In addition,the small batch gradient descent method is used to update the network related parameters during the parameter update process.(2)During the training process,in order to solve the problem of sample input sequence,heuristic learning theory is introduced,so that the sample input during the training process is input in an order from easy to difficult,and the difficulty of the sample depends on the size of the mean square error.At the same time,in order to solve the problem of small sample size in the training process,a penalty term is introduced in the objective function to punish the repetitiveness of the sample input during the training process.At the same time,the strategy of alternating update is adopted to update the relevant parameters in the self-paced regularization term and the weighted loss term.(3)Finally,related experiments and specific index data verify the superiority of the dual-flow neural network compared with other methods under various situations and the error function convergence analysis with the heuristic learning or not. |
