Existing weakly illuminated image enhancement algorithms are strongly dependent on Retinex model and require manual adjustment of parameters. To solve those problems, an algorithm based on Convolutional Neural Network (CNN) was proposed to enhance weakly illuminated image. Firstly, four image enhancement techniques were used to process weakly illuminated image to obtain four derivative images, including contrast limited adaptive histogram equalization derivative image, Gamma correction derivative image, logarithmic correction derivative image and bright channel enhancement derivative image. Then, the weakly illuminated image and its four derivative images were input into CNN. Finally, the enhanced image was output after activation by CNN. The proposed algorithm can directly map the weakly illuminated image to the normal illuminated image in end-to-end way without estimating the illumination map or reflection map according to Retinex model nor adjusting any parameters. The proposed algorithm was compared with Naturalness Preserved Enhancement Algorithm for non-uniform illumination images (NPEA), Low-light image enhancement via Illumination Map Estimation (LIME), LightenNet (LNET), etc. In the experiment on synthetic weakly illuminated images, the average Mean Square Error (MSE), Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity index (SSIM) metrics of the proposed algorithm are superior to comparison algorithms. In the real weakly illuminated images experiment, the average Natural Image Quality Evaluator (NIQE) and entropy metric of the proposed algorithm are the best of all comparison algorithms, and the average contrast gain metric ranks the second among all algorithms. Experimental results show that compared with comparison algorithms, the proposed algorithm has better robustness, and the details of the images enhanced by the proposed algorithm are richer, the contrast is higher, and the visual effect and image quality are better.

For image instance retrieval in the field of computational vision, a semantic region weighted aggregation method based on significance guidance of deep convolution features was proposed. Firstly, a tensor after full convolutional layer of deep convolutional network was extracted as deep feature. A feature saliency map was obtained by using Inverse Document Frequency (IDF) method to weight deep feature, and then it was used as a constraint to guide deep feature channel importance ordering to extract different special semantic region deep feature, which excluded interference from background and noise information. Finally, global average pooling was used to perform feature aggregation, and global feature representation of image was obtained by using Principal Component Analysis (PCA) to reduce the dimension and whitening for distance metric retrieval. The experimental results show that the proposed image retrieval algorithm based on significant semantic region weighting is more accurate and robust than the current mainstream algorithms on four standard databases, because the image feature vector extracted by the proposed algorithm is richer and more discerning.

In order to realize the accurate detection and classification of muscle fatigue states, a new complete muscle fatigue detection and classification system based on human surface ElectroMyoGraphy (sEMG) signals was proposed. Firstly, human sEMG signals were collected through AgCl surface patch electrode and high-precision analog front-end device ADS1299. The time-domain and frequency-domain features of sEMG signals reflecting human muscle fatigue states were extracted after the denoising preprocessing using wavelet transformation. Then, on the basis of the common features such as Integrated ElectroMyoGraphy (IEMG), Root Mean Square (RMS), Median Frequency (MF), Mean Power Frequency (MPF), in order to depict the fatigue states of human muscle more finely, the Band Spectral Entropy (BSE) of frequency domain features of sEMG signals were introduced. In order to compensate the weakness of Fourier transform in dealing with non-stationary signals, the time-frequency feature of the sEMG signals, named mean instantaneous frequency based on Ensemble Empirical Mode Decomposition-Hilbert transform (EEMD-HT), was introduced. Finally, in order to improve the classification accuracy of muscle non-fatigue and fatigue states, the Support Vector Machine optimized by Particle Swarm Optimization algorithm (PSO-SVM) with mutation was used for the classification of sEMG signals to realize the detection of human muscle fatigue states. Fifteen healthy young men were recruited to carry out sEMG signal acquisition experiments, and a sEMG signal database was established to extract features for classification. The experimental results show that, the proposed system can realize high-accuracy sEMG signal acquisition and high-accuracy classification of muscle fatigue states, and its accuracy rate of classification is above 90%.

Due to the limited angle of photography, some natural images are oscured by fence-like occlusion such as barbed wire, fence and glass joints. A novel fence-like occlusion detection algorithm was proposed to repair such images. Firstly, aiming at the drawbacks of the existing fence detection algorithms using single pixel color feature and fixed shape feature, the image was divided into super pixels and a joint feature of color and texture was introduced to describe the super pixel blocks. Thus, the classification of a pixel classification problem was converted to a super pixel classification problem, which inhibited the misclassification caused by local color changes. Secondly, the super-pixel blocks were classified by using the graph cuts algorithm to extend the mesh structure along the smooth edges without being restricted by the fixed shape, which improved the detection accuracy of the special-shaped fence structure and avoided the manual input required by the algorithm proposed by Farid et al. (FARID M S, MAHMOOD A, GRANGETTO M. Image de-fencing framework with hybrid inpainting algorithm. Signal, Image and Video Processing, 2016, 10(7):1193-1201) Then, new joint features were used to train the Support Vector Machine (SVM) classifier and classify all non-classified super-pixel blocks to obtain an accurate fence mask. Finally, the SAIST (Spatially Adaptive Iterative Singular-value Thresholding) inpainting algorithm was used to repair the image. In the experiment, the obtained fence mask retained more detail than that of the algorithm proposed by Farid et al., meanwhile using the same repair algorithm, the image restoration effect was significantly improved. Using the same fence mask, restored images by using the SAIST algorithm are 3.06 and 0.02 higher than that by using the algorithm proposed by Farid et al., respectively, in Peak Signal-to-Noise Rate (PSNR) and Structural SIMilarity (SSIM). The overall repair results were significantly improved compared to the algorithm proposed by Farid et al. and the algorithm proposed by Liu et al. (LIU Y Y, BELKINA T, HAYS J H, et al. Image de-fencing. Proceedings of the 2008 IEEE Conference on Computer Vision and Pattern Recognition. Washington, DC:IEEE Computer Society, 2008:1-8) when using the SAIST inpainting algorithm combined with the proposed fence detection algorithm. The experimental results show that the proposed algorithm improves the detection accuracy of the fence mask, thus yields better fence removed image reconstruction.

Aiming at the problem of impure simplicity and too simple feature extraction of training samples in the existing saliency detection model based on guided learning, an improved algorithm based on Boosting was proposed to detect saliency, which improve the accuracy of the training sample set and improve the way of feature extraction to achieve the improvement of learning effect. Firstly, the coarse sample map was generated from the bottom-up model for saliency detection, and the coarse sample map was quickly and effectively optimized by the cellular automata to establish the reliable Boosting samples. The training samples were set up to mark the original images. Then, the color and texture features were extracted from the training set. Finally, Support Vector Machine (SVM) weak classifiers with different feature and different kernel were used to generate a strong classifier based on Boosting, and the foreground and background of each pixel of the image was classified, and a saliency map was obtained. On the ASD database and the SED1 database, the experimental results show that the proposed algorithm can produce complete clear and salient maps for complex and simple images, with good AUC (Area Under Curve) evaluation value for accuracy-recall curve. Because of its accuracy, the proposed algorithm can be applied in pre-processing stage of computer vision.

Defect detection is an important part of fabric quality control. To make the detection algorithm possess good commonality and high detection accuracy, a dual-scale fabric defect detection algorithm based on sparse coding was proposed. The algorithm combined the advantage of high stability under large-scale and the advantage of high detection sensitivity under small-scale. At first, the dictionaries under large and small scales were obtained through a small-scale over-complete dictionary training method. Then, the projection of detection image block on the over-complete dictionary was used to extract detection characteristics. Finally, the detection results under dual-scale were fused by the means of distance fusion. The algorithm overcame the disadvantage of large computation because of the introduction of dual-scale while using small-scale over-complete dictionary and downsampling the detection image under large-scale. TILDA Textile Texture Data base was used in the experiment. The experimental results show that the algorithm can effectively detect defects on plain, gingham and striped fabric, the comprehensive detection rate achieves 95.9%. And its moderate amount of calculation can satisfy the requirement of industrial real-time detection, so it does have much value of practical application.