A new algorithm of infrared ship wake detection based on fusion of Gabor filter and local information entropy was proposed to solve the problem that the infrared image of ship wake is easily disturbed by the sea clutter,the contrast is low,and the image can not be identified by the traditional method.First of all,the contrast between the wake and the sea background is calculated by using the gray level co-occurrence matrix to determine whether there is a ship wake in the region,and the region of interest is extracted to improve the processing speed of the algorithm.Secondly,multi direction Gabor filter and local information entropy are used for feature fusion to realize the feature enhancement of ship wake.Finally,the infrared ship wake detection is realized by threshold segmentation and Hough transform.Experimental results show that this method can effectively preserve the texture features and details of ship wake,and accurately extract the complete wake edge,which greatly improves detection rate.

The computer can simulate the human visual environment to identify and measure things in the field of vision.With the increase of accuracy,computer vision can replace the function of human’s eyes to achieve simple and repetitive manual operations.The introduction of computer vision into logs can increase the yield of logs,reduce wood loss,maxmize the utilization rate of logs with high-efficiency and accurate performance of the computer,minimize the production of raw materials that generate square waste,and increase the output rate of logs.This algorithm is applied to automated band saw log cutting systems.The basic process includes eliminating image noise through image preproces-sing,removing the background through color segmentation,giving the contour of the region of interest by edge detection,filling the misprocessed contour edges through morphological operations,and calculating the largest area of the fitted ellipse.The experimental results show that the arithmetic can meet the requests of actual production,and the accuracy reaches 95%.

Aiming at the low efficiency of incremental motion recovery structure algorithm in multi-view geometric 3D reconstruction algorithm,a multi-view geometric 3D reconstruction method based on AKAZE algorithm was proposed.The target image obtained by the camera is detected and matched by AKAZE algorithm,and the weak matching image is eliminated by using the random sample consensus algorithm and the three view constraints.Then the global rotation parameters are solved by the least square method according to the relative position and attitude parameters of the matching graphs,and the global displacement parameters are solved by using the three-view constraint relation.Finally,the bundle adjustment optimization is carried out.The experimental results show that the proposed algorithm can improve the processing efficiency and meet the needs of fast processing on the basis of improving the reconstruction effect.

The target template was designed and the BP neural network model was proposed,which can calculate the position deviation parameter between the two cameras according to the image coordinates of the feature points on the target template on the dual camera.GA algorithm is used to optimize BP neural network to compensate the shortco-mings.The training sample data set is used to train the proposed model,and the model is tested with the test sample data set,and finally the training model is used for the actual production of the two-camera module.The actual application results show that the calibration precision and time can meet the requirements of actual production based on the proposed method.

An algorithm of extracting key actions in sign language was brought out in this paper.In the continuous and complex sign language,the number of key actions is small and the state is relatively stable.Thus using the key actions to construct the data model of the sign language will reduce the unstable factors and improve the accuracy.In this paper,an adaptive classification algorithm was proposed,which extracts the key actions step by step according to the time order and the irrelevance among the key actions.Experiments show that the algorithm can be used for the non-specific population.Moreover,the algorithm can extract all the key actions from both the single vocabulary and the continuous sentence.Key actions can be regarded as primitives of sign language,and thus sign language can be looked upon as different combinations of those primitives as well.Therefore,as for the continuous and complex sign language,the extraction of key actions has important significance not only for the data model construction,but also for its recognition.

In order to remove image noise and preserve image edge information in image edge detection,a edge detection method for noisy image based on wavelet transform modulus maxima and improved mathematical morphology edge detection was proposed.Firstly,the image edge detection algorithm based on wavelet transform modulus maxima was introduced.Then a new improved mathematical morphology was proposed.Finally,in order to synthesize the merits of the two algorithms,a new fusion method was used to fuse the results of the two methods together,and a novel edge detection method for noisy image based on wavelet transform and new morphology was proposed.The experimental results show that the proposed fusion detection algorithm can suppress the noise more effectively and improve the edge detection effect than using wavelet transform modulus maxima or new mathematical morphology alone.

Recognizing freehand sketches is a greatly challenging work.Most existing methods treat sketches as traditional texture images with fixed structural ordering and ignore the temporality of sketch.In this paper,a novel sketch recognition method was proposed based on the sequence of sketch.Strokes are divided into groups and their features are fed into recurrent neural network to make use of the temporality.The features from each temporality are combined to produce the final classification results.The proposed algorithm was tested on a benchmark,and the recognition rate is far above other methods.

A hard exudates (HEs) detection method based on deep convolution neural network was proposed in this paper,which achieves the purpose of automatic detection for HEs and contributes to the creation of diabetic retinopathy (DR) computer-aided diagnostic system.This method includes training the classification model for HEs offline and detection for HEs online.In order to train HEs classification model offline,CNN is adopted to extract HEs features automatically.Then,HEs in fundus image are detected by HEs classification model which has been trained offline,meanwhile,HEs probability graph and HEs pseudo-color map are obtained.The method was verified on standard data set and self-built data set respectively.Compared with other methods,the proposed method is profitable with strong robustness,and has very strong clinical practice significance.

In order to realize the safety early warning of dangerous chemical products,this paper used laser ranging and encoder to monitor the five distance in the warehouse.In order to solve the noise problem of ranging data,a new algorithm was designed for reducing noise and feedback compensation.According to the characteristicing of the noise and the distance,taking the tested object as the center,objects are divided into three categories from far to near.The first layer and the second layer use peak denoising,the third layer uses piecewise fitting angle,simultaneously interlayer feedback error correction is used to realize the closed loop denoising.Experiment results show that the variance is reduced by 0.83 compared with the denoising algorithm.Compared with the comparison algorithm of difference value,this algorithm can removed catastrophie caused by a small amount of noise concentrated and the variance value is decreased by 1.93.The algorithm can remove the noise and restore the position of the object.

In order to improve the over-segmentation in the traditional watershed algorithm,a watershed segmentation algorithm of gradient hierarchical reconstruction was proposed under opponent color space,considering the interference of reflected light on the image.Firstly,the color image is converted from RGB space to the opponent color space which has nothing to do with the reflected light.Secondly,the gradient image of the color image is obtained by combining the image information entropy.Thirdly,the gradient image is hierarchically reconstructed according to the distribution information of the gradient histogram.Then morphological minimum calibration technique is used to calibrate the combined gradient image.At last,watershed segmentation is applied to the corrected image.Experiments on different types of images were carried out.The experimental results show that the proposed algorithm is more prominent than the three classic watershed algorithms in the number of divided regions,running time and the DIR.The new algorithm is more in line with human perception of the image,the segmentation and performance are better,and it has higher robustness and practicality.

FXAA is a post-processing anti-aliasing algorithm.Because it is an edge detection algorithm based on image pixel,it causes a lot of unnecessary anti-aliasing computation.In order to improve the performance of anti-aliasing,an improved anti-aliasing algorithm based on FXAA (IAAFXAA) was proposed.The depth and normal of the relative view are saved into the texture.The algorithm extracts depth and normal information from G-buffer,and uses depth and normal information to perform more accurate edge detection.A large number of experimental results and analysis show that while ensuring the good anti-aliasing effect,the proposed algorithm can determine the anti-aliasing region more accurately to generate high-quality boundaries,and avoid excessive blurring of images to improve image quality.

An improved ORB feature extraction algorithm based on quadtree encoding was proposed in this paper,which can solve the problem that the detected feature points are too dense to show the picture information completely.Firstly,the image pyramid is built to make the scale invariance.Then,the feature points are extracted on each image pyramid and quadtree encoding is introduced to homogenize the feature point.Finally,the direction and descriptor are calculated for each feature points.In this paper,the Xtion PRO was used as an experimental tool to extract the feature points under indoor environment,and the proposed algorithm was compared with others.Experimental results show the effectiveness and accuracy of the proposed method.

In the current integrated agricultural pest control,agricultural insect pests are detected primarily by professionals’ sample collection and sorting manually,such manual classification method is both expensive and time consuming.Existing computer-aided automatic detection of agricultural pests has a high requirement on the background environment of pests and cannot locate agricultural pests.To solve these problems,this paper proposed a new method for automatic detection of agricultural pests based on the idea of the deep learning.It contains the region proposal network and the Fast R-CNN network.Region proposal network extracts feature in one or more regions of arbitrary size and complicated background images,then gets preliminary position of the candidate regions of agricultural pests.Preliminary position of the candidate regions of agricultural pests is taken as an input to Fast R-CNN.Fast R-CNN finally learns the classification of target in the preliminary location candidate area and calculates exact coordinates by studying the intraspecific differences and interspecies similarity of agricultural pests.Meanwhile,this paper also established a labeled actual scene tag agricultural pests database,and the proposed method was tested on this database,with theaverage precision up to 82.13%.The experimental results show that the proposed method can effectively enhance the accuracy of agricultural pests detection,and get accurate positions,and is superior to the previous automated agricultural pest detection methods.

Visual tracking is an important part of the computer vision,and kernelized correlation filter tracking is a relatively novel method in visual tracking field.It is different from traditional method based on target feature,which has high accuracy and fast tracking speed.However,when the object moves rapidly or has the larger scale changes,the method cannot track the target accurately.This paper proposed an improved algorithm based on the correlative filter which can effectively overcome the above problems.The learning factors of kernelized correlation filtering and the ada-ptive updating model of learning factors are determined by using random update multi-template matching.Experimental results show that the algorithm can adjust the learning factors quickly according to different scenarios,thus the success rate of tracking will be improved.Through adaptive learning factor and multi-template matching,this algorithm has robust adaptability to partial occlusion,illumination and target scale.

A gaussian process assisted covariance matrix adaptation evolution strategy(GPACMA-ES)optimization algorithm was proposed in this paper.The kernel function used in the GPACMA-ES algorithm is constructed by the cova-riance matrix.Taking advantage of the Gaussian process,which plays a key role in both online learning about the histo-ric experience and predicting the promising region which contains globally optimal solution,the frequency of calculating fitness function in the algorithm is reduced markedly.Meanwhile,in order to improve the efficiency of the algorithm,GPACMA-ES is sampling in the trust region.So it has rapid convergence and good global search capacity.Finally,a case study of medical image registration is examined to demonstrate the ability and applicability of the GPACMA-ES.Expe-riment results show that GPACMA-ES is proper for medical image registration than CMA-ES,and it has a better effect on the precision of registration while reducing the number of calculation of the fitness function.

The input of traditional artificial neural network is in vector form,but the image is represented by matrix.Therefore,in the process of image processing,the image will be inputted into the neural network in vector form,which will destroy the structure information of image,and thus affect the effect of image processing.In order to improve the ability of network on image processing,the multilayer feedforward neural networks with matrix inputs are introducedbased on the idea and method of deep learning.At the same time,the traditional back-propagation algorithm (BP) is used to train the network,and the training process and training algorithm are given.After a lot of experiments,the network structure with good performance were determined,and the numerical experiments were carried out on the USPS handwritten digital data set.The experimental results show that the proposed multilayer network has better classification results than the single hidden layer feed forward neural network with matrix input (2D-BP).In addition,to deal with the problem of color image classification,this paper provided an effective and feasible method,the new 2D-BP network,to deal with it

This paper proposed an image feature description method based on complex network.By using the key points of the image as the node of complex network,this method uses MST measure to achieve dynamic evolution process,anduse complex network characters in different phase to achieve the description of the shape of the image.With the distance and the difference of gray level between a pixel and its neighborhood,a series of degree matrices can be represented by using a series of thresholds,and the texture feature can be represented by calculating the degree distribution of network nodes under different thresholds.This method is based on statistical image description method.It has stronger robustness and rotation invariance,and has a great performance in classification experiments.

There are some problems with the color image segmentation technology based on traditional Fuzzy C-means clustering algorithm,such as the selection of the initial category number,the determinated of the initial centroids,large amount of calculation in clustering process and post-processing.Based on the research of these problems,according to the shortage of random initialization in traditional FCM,and for getting more accurate initialization automatically,this paper proposed a clustering segmentation method combining Hill-climbing for color image(HFCM),which can generate the initial centroids and the number of clusters adaptively according to the three dimensional histogram of the image.In addition,a new post-processing strategy which combined the most frequency filter and region mergeing was introduced to effectively eliminate small spatial regions.Experiments show that the proposed segmentation algorithm achieves high computational speed,and its segmentation results are close to human perceptions.

In order to make full use of the color,shape,texture and other features in the 3D model,a 3D model retrieval method was proposed based on angle structure features of render images.Firstly,the 3D model render images are taken as a test dataset and the marked natural images are taken as a training set.The render images are classified based on their skeleton-associated shape context and the angle structure features are extracted to establish the feature library.Then,the angle structure features of the input natural images are extracted.The distance measurement method is used to calculate the similarity between the angle structure feature of input natural image and those features in the feature library.The experimental results show that the full utilization of the color,shape and color space information of the render image is an effective way to achieve 3D model retrieval.

This paper proposed a novel double level set algorithm based on NL-Means denosing method for brain MR image segmentation,which has a large amount of noise and complicated background,and cannot be separated completely by traditional level set.First of all,this algorithm gets the denoised image by analyzing the image with NL-Means denosing method.Then,the algorithm identifies denoised image by segmenting the analyzed results in terms of improved double level set model.In order to deal with the effect of intensity inhomogeneities on the medical image,the algorithm introduces a bias fitting term into the improved double level set model and optimizes the denosing method result.The experimental result shows that the algorithm can reduce the problems of intensity inhomogeneities and noise,can separate brain MR image including intensity inhomogeneities and noise completely,and can obtain the expected effect of segmentation.

In the field of image retrieval,in order to make the image retrieval more convenient and efficient,this paper proposed a new image retrieval feature,namly local autocorrelation feature,which provides a new tool for content-based image retrieval.It has the characteristics of orientation feature and texture feature.The experiment was carried out for local autocorrelation feature presented in this paper on the Corel10K database,the experimental results show that the average retrieval precision and recall rate of the local autocorrelation feature are lower than the color feature,but it is higher than that of the orientation feature.In addition to color features,the local autocorrelation feature is an efficient image retrieval feature.

Due to that CNN can effectively extract deep features of the image,this paper used GoogLeNet and AlexNet models which have excellent performance in image classification to diagnose the pathological image of gastric cancer.Firstly,according to the characteristics of medical pathological images,this paper optimized the GoogLeNet model to reduce the computational cost under the premise of ensuring the accuracy of diagnosis.On this basis,it proposed the idea of model fusion.By combining more images with different structures and different depths,more effective pathological information of gastric cancer can be acquired.The experimental results show that the fusion model with multiple structures has achieved better results than the original model in the diagnosis of pathological images for gastric cancer.

Single tree detection can assist forestry statistics in getting information such as position,width and diameter of the crowns,so it is of great significance for the development of precision forestry.In order to solve the problem of inaccurate canopy delineation in single-tree canopy detection,this paper proposed a single tree detection algorithm based on morphological Snake model for remote sensing images.Firstly,the forest features are analyzed.Then the local maximum method is used to extract treetops according to the forest feature map and the distance map.After this,the contour of Snake model is initialized for all crowns according to treetops,after evolution of the contour,the final detection result of individual trees is obtained.In order to verify the effectiveness of the method,this paper gave comparative analysis of the region growing method,template matching method,watershed method and the proposed morphological snake model method.The experimental results show that the proposed method is more accurate and the shape of the crown is more realistic.Compared with the other three methods,the detection score is 6% higher and the area average difference is reduced by 0.5m².

In order to meet the actual demand,a new detection method of steel rail abrasion based on line laser image processing was proposed after analyzing current methods and characteristics of steel rail abrasion detecting equipment at home and abroad.The bending degree of line laser image on the wear of rail was used to determine the width and depth of steel rail abrasion.The edge points and centre points could be found by using roof-type edge detection method,then straight lines can be fitted by using these points.The optimal features combination is selected by removing the redundant features with high correlation.Finally,the experiment results show that the method could extract features amount effectively,and obtain the width and depth of the steel rail abrasion accurately.The characteristics of algorithm inculde small amount,simple and high precision.It lays the foundation for the development of steel rail abrasion detection device.

Principal component analysis (PCA) is one of the most important data reduction algorithms,there is much-maligned views in the process of handling data.A novel improved similar principal component analysis (SPCA) algorithm which is based on principal component analysis (PCA) algorithm was proposed in this paper.This algorithm can keep some detail information in the process.Taking the MNIST handwritten numeral database as an example, the near feature vector is chosen in original vectors to get the groups of non-orthogonal feature vectors.Then,the vectors of trai-ning library is compared with the vectors of testing library,and the recognition rate is calculated.Recognition results indicate that the algorithm can make high identification of the testing samples through a small number of training samples.