To efficiently improve the effects of image Super Resolution (SR), a multi-stage cascade residual convolution neural network model was proposed. Firstly, two-stage SR image reconstruction method was used to reconstruct the 2-times SR image and then reconstruct the 4-times SR image; secondly, residual layer and jump layer were used to predict the texture information of the high resolution space in the first and second stages, and deconvolution layer was used to reconstruct 2-times and 4-times SR images. Finally, two multi-task loss functions were constructed respectively by the results of two stages. And the loss of the first stage guided that of the second one, which accelerated the network training and enhanced the supervision and guidance of the network learning. The experimental results show that compared with bilinear algorithm, bicubic algorithm, Super Resolution using Convolutional Neural Network (SRCNN) algorithm and Fast Super Resolution Convolutional Neural Network (FSRCNN) algorithm, the proposed model can better construct the details and texture of images, which avoids the image over smoothing after iterating, and achieves higher Peak Signal-to-Noise Ratio (PSNR) and Mean Structural SIMilarity (MSSIM).

To improve the recognition rate and reduce the false-recognition rate in real-time detection of flame in video surveillance, a fast flame recognition algorithm based on multi-feature logarithm regression model was proposed. Firstly, the image was segmented according to the chromaticity of the flame, and the Candidate Fire Region (CFR) was obtained by subtracting the moving target image with reference image. Secondly the features of the CRF such as area change rate, circularity, number of sharp corners and centroid displacement were extracted to establish the logarithmic regression model. Then, a total of 300 images including flame and non-flame images, which were got from National Institute of Standards and Technology (NIST), Computer Vision laboratory of Inha University (ICV), Fire detection based on computer Vision (VisiFire) and the experimental library consisting of the candle and paper combustion were used to parametric learning. Finally, 8 video clips including 11071 images were used to validate the proposed algorithm. The experimental results show that the True Positive Rate (TPR) and True Negative Rate (TNR) of the proposed algorithm are 93% and 98% respectively. The average time of identification is 0.058 s/frame. Because of its fast identification and high recognition rate, the proposed algorithm can be applied in embedded real-time flame image recognition.

Aiming at the problem of small gradient, low learning rate, slow convergence of error during the process of using Deep Belief Network (DBN) algorithm to correct connection weight and bias of network by the method of back propagation, a new algorithm called Multi-Innovation DBN (MI-DBN) was proposed based on combination of standard DBN algorithm with multi-innovation theory. The back propagation process in standard DBN algorithm was remodeled to make full use of multiple innovations in previous cycles, while the original algorithm can only use single innovation. Thus, the convergence rate of error was significantly increased. MI-DBN algorithm and other representative classifiers were compared through experiments of datasets classification. Experimental results show that MI-DBN algorithm has a faster convergence rate than other sorting algorithms; especially when identifying MNIST and Caltech101 dataset, MI-DBN algorithm has the fewest inaccuracies among all the algorithms.

An automatic segmentation method based on image patch matching strategy was proposed to realize the automatic segmentation of glomerular basement membrane automatically. First of all, according to the characteristics of the glomerular basement membrane, the search range was extended from a reference image to multiple reference images, and an improved searching method was adopted to improve matching efficiency. Then,the optimal patches were searched out and the label image patches corresponding to the optimal patches were extracted, which were weighted by matching similarity. Finally, the weighted label patches were rearranged as the initial segmentation of glomerular basement membrane, from which the final segmentation could be obtained after morphological processing. On the glomerular Transmission Electron Microscopy (TEM) dataset, the Jaccard coefficient is between 83% and 95%. The experimental results show that the proposed method can achieve higher accuracy.

An improved Principal Component Analysis (PCA) method was proposed with the synchronous multi-dimensional data stream anomaly analysis techniques. In this method, the problem of the original data stream variation tendency was mapped to the eigenvector space, and the steady-state eigenvector was solved, then the abnormal changes of the synchronous multi-dimensional data stream could be diagnosed by the relationship between the instantaneous eigenvector and the steady-state eigenvector. This method was applied to the abnormality diagnosis of the tunnel strain monitoring data stream, and the real-time monitoring and warning system for the tunnel strain was realized by using VC++. The experimental results show that the proposed method can reflect the changes of the aperiodic variables timely and realize the anomaly monitoring and early warning for multi-dimensional data stream effectively.

In order to speed up the convergence of self-training AM-CNN (Associative Memories Cellular Neural Network) and enhance the performance of achieved AM-CNN, an algorithm for obtaining the space-invariant cloning templates of AM-CNN was proposed, which took the output error of objective CNN as objective function and took local searching and global searching respectively in two internals separated by a given objective function threshold, coupled with the idea of ant optimization algorithm and Particle Swarm Optimization (PSO). Concluded from the numerical simulation results, the proposed algorithm outputs the objective AM-CNN and converges quickly. Meanwhile, the performance of the achieved AM-CNN is better and more stable compared with previous methods. The achieved AM-CNN is also robust to Gauss noise of N(0,0.8) with recall rate of about 90%.