Object detection is one of the core tasks in the field of computer vision. In recent years, with the rapid development of deep learning, the object detection technology based on deep learning has become the very popular mainstream algorithm. It has been widely used in many fields, such as face detection, vehicle detection, pedestrian detection, and unmanned driving, etc.. This paper systematically summarizes the current research progress of deep learning-based object detection algorithms, and thoroughly analyzes the advantages and disadvantages of each algorithm and its results on the datasets VOC2007 and COCO. Finally, the future development of object detection based on deep learning is also discussed in this paper.

Due to the presence of background noise in natural scenes and the interference of complex factors such as illumination, rotation, and shooting angle, it is very difficult to identify the image of buildings in natural scenes. Aiming at the dependence of traditional building extraction methods on human design and the improvement of building edge feature extraction algorithm.Through the Keras framework to obtain the bottleneck layer of convolutional neural networks (CNN) model MobileNet,and add a new classifier for transfer learning. A large number of data augmentation and test set augmentation are applied to the input image. After three versions of transfer learning, high accuracy was achieved within 480 iterations in three test set. Compared with other feature extraction algorithms, CNN has the advantages of non-transformation and automatic extraction of features, achieves higher accuracy in a shorter period of time. At the same time, MobileNet weight only occupy 15.3 MB with high precision and less calculation, which can be widely transplanted to mobile devices. The system based on model migration has the functions of photo recognition, photo album recognition, menu display, etc., providing mobile platform users with a convenient and simple tool to quickly and accurately obtain the information of buildings in natural scenes.

At present, the trajectory anomaly detection algorithm is mainly based on the trajectory space shape, ignoring the internal feature information of the trajectory. In this paper, based on the characteristics of trajectory structure, the internal and overall characteristics of trajectory are analyzed, and a trajectory structure anomaly detection method based on feature entropy is proposed. Firstly, the trajectory is divided into trajectory segments according to the opening angle, and the local features of the trajectory segments are fitted by linear regression model to complete the trajectory segment division. Secondly, the trajectory structure framework is introduced to describe the internal characteristic attributes of the trajectory, and the trajectory structure distance is used to measure the distance between trajectory segments. Meanwhile, the method of feature assignment based on entropy is proposed, which comprehensively considers the influence of the internal characteristics of the trajectory. Then, DBSCAN clustering algorithm is used to divide the trajectory set into several clusters and extract representative trajectories. Finally, by comparing the structural similarity between the trajectory segments and the representative trajectories, the abnormal trajectory segments are extracted, and then the abnormal trajectories are excavated as a whole. Experiments using multiple data sets show that the trajectory structure anomaly detection method based on feature entropy can detect information anomaly from the trajectory spatial shape and internal feature attributes. Obvious abnormal trajectories and their segments can be found in an all-round way, which makes the detection results more meaningful.

The traditional co-clustering algorithm simultaneously produces a predetermined number of partitions for rows and columns of a two-dimensional data matrix. Most existing co-clustering algorithms are designed for non-overlapping and exhaustive co-clustering. However, many real-world datasets contain not only a large amount of overlap between row and column clusters, but also outliers that do not belong to any cluster. In view of this, an overlapping co-clustering algorithm is proposed by maximizing modularity (OMMCC), that is, both row clusters and column clusters are allowed to overlap, and the row and column outliers of the data matrix are not assigned to any cluster. Specifically, a unified framework is designed to add non-exhaustive and overlapping constraints to the objective function. Through using an iterative alternating optimization process to directly maximize the modularity, the better block diagonal non-exhaustive overlapping co-clustering can be obtained efficiently. Besides, the degree of overlap and non-exhaustive parameters are easy to understand. The experimental results show that the proposed method is very effective, stable and superior to other co-clustering algorithms.

Image-to-image translation is a class of tasks which translate an image to another image of the specified type. In essence, it is a mapping problem from pixels to pixels.However, existing methods have showed limitation in term of scalability and robustness when it comes to translation tasks for more than two domains.In order to achieve translation results of high quality and high efficiency, an unpaired image-to-image translation based on conditional projection is proposed in this paper. The proposed method calculates the similarity between the feature information learned by generator and conditional information, which can improve the accuracy of translation and qualities of generated images. Compared to existing models, the proposed method adopts fewer parameters and shorter training time. The effectiveness of the proposed model is shown on multiple datasets.

Aimed at the problem of lacking of emotions images will influence the performance of CNN seriously, a large image emotion dataset, Large-scale deep emotion (LSDE) dataset, is built using semi-supervised dynamic method to ensure emotion labels of images. In order to bridge the gap and find the relationship between different levels features effectiveness, at first, objects of images are divided from complete images using salient object detection, then a relationship learning network is adopted based on foreground images and background images in this paper to learn relationships between different levels features of images to bridge the gap between image features and emotions of human beings. Experimental results in LSDE dataset, Twitter2 dataset and ArtPhoto dataset show that relationship learning network can extract different level features from foreground and background images and can learn the relationship between different level features and bridge the gap between features of images and emotions of human beings, the image emotion recognition system also can recognize emotions of images accurately.

Tuning the weights of deep neural networks using loss and back propagation algorithm has been widely used in image retrieval. Applying triplet ranking loss to tune the weights can make the generated image representations preserve more semantic features. However, the relations among different categories of images are not fully considered in the triplet ranking loss.The quadruplet complete loss is proposed based on that inter-class similarity is smaller than the intra-class similarity, and the similarities among the query image and similar or dissimilar images are also fully considered in the loss.Further more, an effective quadruplet based deep hashing network architecture is also proposed for image retrieval. The experimental results show that our method can achieve excellent retrieval performance in CIFAR-10, SVHN and NUS-WIDE.

The knowledge discovery of multi-source data has been one of the most important research spots in Big Data era, however, relevant researches in formal concept analysis are few. Based on which, multi-source formal decision context is firstly defined in this paper. Then, the granular attribute reduction of multi-source formal decision contexts is studied, and the connections with the granular attribute reduction of common formal decision contexts are also researched. Finally, the granular rules acquisition method of multi-source decision formal contexts is proposed.

In the decision evolution set theory system, the forecast rules are accompanied by the real rules, so the forecast rules will inevitably have an impact on the real rules. However, there are few studies on the relationship between forecast rules and real rules. In this paper, the game theory method is introduced to construct the game matrix of forecast rules and real rules, calculate their profits, and then analyze their influence on the decision information system.

Feature selection is an essential step for analyzing gene expression datasets with very much high dimensions and small number of samples. However, the available feature subset selection algorithms share the common deficiencies that the feature subset is dependent on the training subset, and is various with different training samples. In order to solve this problem in feature selection, a new ensemble feature selection algorithm based on the kernel extreme learning machines is put forward. 5-fold cross validation experiments are adopted to partition the original dataset. For each training subset, 5-fold cross validation experiments are adopted again to partition it, then feature selection process has been done on each sub-training subset, and the union of the five selected feature subsets constructs the feature subset corresponding to the training subset. The classification power of the feature subset is evaluated by the performance of the kernel extreme learning machine built on it. The stability of feature subsets detected by the feature selection algorithms is evaluated by the mean Jaccard coefficient of five feature subsets obtained by 5-fold cross validation experiments on original data. The performance of the proposed ensemble feature selection algorithm is tested on five gene expression datasets. The performance of the proposed feature selection algorithm is compared to the available ones, including SVM-RFE, LLE Score, ARCO, DRJMIM, Random Forest, and mRMR. All the experimental results show that the proposed ensemble feature selection algorithm can not only detect the stable feature subset, but also can select the feature subset with high predictive power.

In supervised classification learning, the impact of label noise on the model is often more important. The existing label noise filtering methods generally detect and remove noise samples based on the prediction results of the model. When the number of noise samples is large, removing the noise samples will affect the integrity of the original samples and make the sample information missing. Aiming at this problem, a method of label noise cleaning based on active learning is proposed, namely GP_ALNC(active label noise cleaning based on classification with Gaussian process). This method combines Gaussian process model and active learning to select the most uncertain samples from existing labeled sample sets and outsourcing them to artificial experts for examining. The proposed iterative method can clean away most of the noise data while maintaining the integrity of the original data. For the label noise problem in the two-class task, the proposed method is compared with the existing methods ALNR(active label noise removal) and ICCN_SMO(iterative correction of class noise based on SMO) on the MNIST and UCI data sets. The experiment results show that the proposed GP_ALNC may achieve good performance.

Aiming at the defect of distinguishing similar motion from the traditional motion history image, a behavior recognition method combining improved motion history image and support vector machine was proposed. Firstly, the moving target of video frame was extracted and the external rectangular box of the moving target was marked out, the optical flow vector was calculated for each pixel in the rectangular region. Secondly, the gray value of each foreground pixel in the motion history image was set as the sum of the optical flow length at the pixel position and the historical gray value of a certain weight. While the gray value of each background pixel was attenuated by weight. Finally, the Hu moments were extracted from the motion history images and were sent as input to the SVM classifier for classification, thus completing human behavior recognition. The experimental results on KTH dataset show that the proposed method can meet the real time requirement and the recognition rate can reach 99%.

Aiming at severe weather conditions, traffic flow predition modd based on random forest was proposed.Based on taxi data and weather conditions in New York city in 2016, screening the original GPS data layer by layer, the data that meet the definition of severe weather conditions are screened out. Based on the random forest regression method, the traffic flow prediction model under severe weather is studied, and the performance of the model is improved by adjusting the super parameters of the model. At the same time, the performance of random forest model is compared with that of BP neural network model and decision tree model, and the experimental results of random forest prediction model are better.

In order to avoid overfitting in deep convolutional neural networks, a new regularization method called deterministic DropConnect is proposed. The main idea of the deterministic DropConnect is to reduce the spatial dimension of the convolution filter by selectively discarding the partial connection, according to the contributions of the different weights of the convolutional filters to the result. The connections between the layers of the convolutional neural networks are more sparse. In this paper, the effectiveness of the proposed deterministic DropConnect is evaluated on the task of image classification. The error rates are 0.32%, 5.33% and 26.88% in the MNIST, CIFAR-10 and CIFAR-100 datasets, respectively. Compared with the original experimental error rates, the error rates reduce by 0.15%, 1.09% and 1.36%, respectively. Experiments show that the proposed algorithm is an effective method to deal with the overfitting in neural networks, and can improve the robustness and generalization ability of the convolutional neural networks.

The state-of-the-art total variability factor analysis in language recognition can only preserve the global structure of speech data, without mining the local structure information, and the language of training speech data is not considered. To solve the problem, neighborhood preserving embedding (NPE) algorithm is introduced into the language recognition system. NPE can preserve the local neighborhood structure of speech data by constructing a neighborhood graph. As well，NPE can effectively use the language label information of speech data by supervised training. The proposed method is compared with the total variability factor analysis in 30 s and 10 s tasks of the NIST 2011 language recognition evaluation (LRE) dataset. The experimental results indicate that the proposed NPE method can overcome the deficiency of total variability factor analysis and improve the system performance significantly.

Graph model reasoning is an important work in graph model research. How to verify the designed reasoning algorithm needs to be tested based on a large number of experimental samples. In order to verify the effectiveness and solution time of the optimal query algorithm, the experimental data designed is critical. In this paper, the random graph model is designed according to the structural characteristics and parameter characteristics of the graph model. The algorithm designed in this paper generates CP-nets of random structure according to the number and degree of vertices. The principle is getting DAG code by improving Prüfer code, furthermore, the one-to-one mapping between DAG code and graph structure is established to realize the random generation of graph model. Then, by combining the designed dominant query algorithm with the typical dominant query. It is proved that the graph model generated by the designed graph model generation algorithm has the randomness of corresponding characteristics. The time consumption of the dominant detection algorithm is heavily dependent on the randomness of topological structure of the graph and the number of parameters.

In order to process the image with for high correlation, high redundancy and enormous data, a new image encryption method has been devised based on true random number and pseudo random number. At first, initial conditions of row scrambling and column scrambling are calculated. Two-dimensional Logistic map is applied to generate chaotic sequences. Extended XOR is executed to enhance system security. Then, row scrambling and column scrambling are performed to improve system security. Finally, image diffusion is applied and encryption image is obtained. Experiment results and theoretical analysis prove that the proposed method is effective and can resist many kinds of attacks.

At present, image privacy-preserving is mainly applied to the field of cloud computing, and since the recognition tasks for images or videos generally need to be visually visible, the privacy-preserving problem is often ignored. To solve this kind of problem, inspired by sparse representation for classification based on compressed sensing (CS) which is robust at the images with occlusions and disguises, an extended model of single-layer CS sampling is proposed to ensure the image is degraded and the content is gradually indiscernible after multilayer CS sampling and encoding, which can still be used for image recognition and achieve the purpose of privacy-preserving. To be able to effectively evaluate image content privacy-preserving for multilayer CS sampling and coding, a content privacy-preserving evaluation (MCS-CPPE) model for multilayer CS images based on human visual system (HVS) is proposed, which by measuring image contrast and extracting local binary pattern (LBP) feature because of degradation of contrast and image visual structure. Experiments with human visual correlation on the three constructed databases show that the proposed model has a better prediction performance and effect.

With the rapid development of deep neural networks, model compression technology has become indispensable in the process of making deep learning models reliably deployed on embedded systems with limited resources. At the same time, exploring the adversarial robustness of neural networks has recently gained more and more attention, because recent works have shown that these models are susceptible to adversarial attacks. Model compression and robustness play an important role in the deep learning model from landing to practical application scenarios. However, in the existing literature, the two have been mostly studied independently, so this paper aims to combining model compression and robustness to make the model compact and robust concurrently. In the framework of adversarial training, we have studied some of the properties of the relationship between model compression and model robustness. And it is proved by experiments that the model compression and the anti-bracket robustness can be obtained simultaneously.

Traditional centrality based methods ignore the influence of the loop structure in the process of dismantling the network,and traditional decycling based methods mistakenly delete a large number of non-critical nodes in the process of decycling. To solve these problems, a network dismantling method based on neighbor nodes fusions is proposed, which takes the influence of loop structure into account on network dismantling and reduces the removing of non-critical nodes effectively through neighbor nodes fusion. Meanwhile, the method further reduces the deletion of non-critical nodes by using a node put-back strategy. The experimental results show that the proposed method can accurately select the critical nodes of network dismantling and fully dismantle the network with fewer critical nodes deletion. It shows stable performance and strong adaptability with various network structures.

In order to improve the accuracy of the recommendation system to recommend new products to users, it is necessary to discover the hidden preferences of each customer and the performance of each product. User feedback techniques were often used to discover potential characteristics and user dimensions of a product.A recommendation model is presented that combines potential factors in user ratings with potential topics in reviews. The model enables more accurate scoring predictions by utilizing the information presented in the review text, and is particularly useful for scoring predictions for new products and new users. Through the verification experiments on the public dataset, it is proved that the model has a significant improvement in performance compared with the traditional recommendation system.

COVID-19 has severely been threatening human being. The computer aided automatic segmentation of lung CT (computed tomography) images is an important way to help medicine doctors to make fast and accurate diagnostic decisions. Therefore, a lightweight model referred to as COVIDSeg is proposed in this paper for segmenting the lung CT images of COVID-19 patients. This model adopts encoder-decoder structure. The squeeze and extend channel attention block (SECA) and the residual multi-scale channel attention block (RMSCA) are proposed to compose the main components of the encoder subnetwork. The dual-path structure is proposed to connect each module of the encoder network. The features are transferred layer by layer along the path, and features from different layers interact between pathways, so that the feature information can be transferred and expressed between different layers. The feature aggregation module is used as the main component of the decoder network, so as to realize the multipath decoder via multiscale feature decoding. This COVIDSeg model is tested on four public COVID-19 CT image datasets. The experimental results show that the proposed lightweight model COVIDSeg for segmenting the COVID-19 CT images outperforms the current main medical image segmentation models in terms of several popular metrics, and it is so far the best model for segmenting the lung CT images of COVID-19 patients. Furthermore, the ablation experiment was carried out to test the influence over the COVIDSeg model performance from its main modules. The experimental results show that the proposed SECA module and RMSCA module are effective to advance the performance of the COVIDSeg model.The segmentation results are visualized for this COVIDSeg model for the lung CT images of COVID-19 patients, and it is found that the segmentation results are basically identical to the real Mask of the images.

In the static grid map environment, aiming at the problems of traditional ant colony algorithm in robot global path planning, such as failing to find the shortest path, blind path search and many inflection points, an improved ant colony algorithm was proposed. On the basis of traditional forward search, reverse path search was added, that was, bidirectional path search. Different heuristic functions were used in each direction. The combination of the two increased the purpose of path search and the construction efficiency of solution.Path crossing for each generation of paths was helpful to generate new solutions and avoid falling into local optimization.The pheromone of the obtained effective path was updated to avoid the interference of the invalid path.The improved volatilization coefficient formula was adopted to dynamically adjust the volatilization coefficient and set the pheromone concentration range, which could avoid falling into premature maturity.The improved Bessel curve was used to optimize the above optimal path, and the degree of curve optimization could be adjusted according to the demand.Compared with other algorithms, the feasibility, effectiveness and superiority of the improved algorithm are verified.

There exists serveral challenges inMRI images of bladder cancer, such as unclear tumor boundaries, small tumor areas, and discontinuous tumor distribution.The existing segmentation algorithms have huge parameters and complex calculations, and the accuracy and efficiency of the existing methods for bladder tumor segmentation need to be improved. In responseto the above issues, a lightweight bladder cancer segmentation algorithm based on multi-scale feature fusion (PylNet) was proposed in this paper. The algorithm can extract information of different scales through the multi-scale semantic feature extraction module. This module was designed in the coding stage to ensure the reliability and comprehensiveness of the extraction of information on small tumor regions. At the same time, the fusion module designed in another stage can quickly complete the tumor region segmentation, and the amount of parameters used is much less. Experimental results show that compared with FCN8s, SegNet, U-Net and other algorithms, the segmentation accuracy of this algorithm is improved to a certain extentwhere DSC reaches 88.40%, and the parameter amount is 1/13 of FCN8s, which a fast bladder MRI segmentation is achieved.

Smoking behavior detection in public places has become the focus of social attention. At present, smoking behavior detection based on machine vision mainly uses deep learning to detect cigarette ends and decide whether there is a smoking behavior by whether there are cigarette ends. Because the cigarette end target is small, the accuracy of this method is relatively low.Through a large number of observations, it is found that smoking behavior is rhythmic and cyclical. Therefore,a smoking behavior detection method based on human keypoints and YOLOv4 is proposed.On the basis of deep learning to detect cigarette butts, the keypoints detection of human body is added to judge whether smoking action occurs by calculating distance, angle and time cycle. Firstly, AlphaPose and RetinaFace are used to obtain the position information of key points of human body.Secondly, based on the position information of keypoints, the distance between hand and mouth, the angle between hand, elbow and shoulder, and the time period of smoking are calculated, and the smoking behavior rules are set.Finally, the cigarette end in the image is detected with YOLOv4 to determine whether there is smoking behavior. The experimental results show that the smoking behavior is detected timely and effectively in the self collected smoking data.

In order to verify whether the context environment in which users rate items would affect their preferences, according to the Markovian factorization of matrix processes method, a streaming recommendation algorithm based on context (C-SRA) is proposed. On the basis of this method, the selected context information is divided into subjective context and objective context, and the two types of context information and algorithm are integrated in turn. Finally, two groups of experiments based on LDOS-CoMoDa data set show that the C-SRA performed better than the other comparison algorithms in both rate prediction and the recommendation.

An end-to-end method based on sequence labeling while making full use of ontology knowledge to enhance the ability of named entity recognition to complete dialogue state tracking is proposed. On the one hand, this method uses a named entity recognition pointer to label the ontology knowledge information contained in the dialogue history based on the sequence labeling method and effectively uses the slot value to enhance the named entity recognition ability of the ontology set. On the other hand, the pointer network is used to retain the ability of the new slot value recognition. The experimental results show that the method proposed in this paper improves the ability of named entity recognition by 1.2% compared with the existing model, and retains the advantages of new slot recognition scalability.