Explore the vast range of titles available.

You only look once (YOLO) is being used as the most popular object detection software in many intelligent video applications due to its ease of use and high object detection precision. In addition, in recent years, various intelligent vision systems based on high-performance embedded systems are being developed. Nevertheless, the YOLO still requires high-end hardware for successful real-time object detection. In this paper, we first discuss real-time object detection service of the YOLO on AI embedded systems with resource constraints. In particular, we point out the problems related to real-time processing in YOLO object detection associated with network cameras, and then propose a novel YOLO architecture with adaptive frame control (AFC) that can efficiently cope with these problems. Through various experiments, we show that the proposed AFC can maintain the high precision and convenience of YOLO, and provide real-time object detection service by minimizing total service delay, which remains a limitation of the pure YOLO.

We propose a real-time dehazing algorithm for hazy images and videos based on multiscale guided filtering. The most time-consuming step in physical model-based algorithms is estimating the transmission map and atmospheric light. In this work, we develop a computationally efficient approach for the estimation. First, we construct an image pyramid from a hazy image. Then, we estimate the transmission map and atmospheric light at the coarsest level. Next, we obtain the transmission at the finest level by iterative upsampling with guide image filtering to avoid information loss. Furthermore, we extend the single-image dehazing algorithm to real-time video dehazing to reduce flickering artifacts in dehazed videos by making transmission values temporally coherent. Experimental results show that the proposed algorithm is applicable in real-time applications, while providing comparable or even better performance than that of state-of-the-art algorithms.

Achieving a satisfactory cancer classification accuracy with the complete set of genes remains a great challenge, due to the high dimensions, small sample size, and presence of noise in gene expression data. Feature reduction is critical and sensitive in the classification task, most importantly in heterogeneous multimedia data. One of the major drawbacks in cancer study is recognizing informative genes from thousands of available genes in microarray data. Traditional feature selection algorithms have failed to scale on large space data like microarray data. Therefore, an effective feature selection algorithm is required to explore the most significant subset of genes by removing non-predictive genes from the dataset without compromising the accuracy of the classification algorithm. The study proposed an information Gain – Modified Bat Algorithm (InfoGain-MBA) features selection model for selecting relevant and informative features from high dimensional Microarray cancer datasets and evaluate the approach with four classifiers - C4.5, Decision Tree, Random Forest and classification and regression tree (CART). The results obtained show that the proposed approach is promising for the classification of microarray cancer data. The random forest has 100% accuracy with few genes in all seven datasets used. Further investigations were also conducted to determine the optimal threshold for each of the datasets.

In this paper, a novel technique for image classification is proposed with the three main contributions. First, we give the texture extraction technique for each image to have the two-dimensional interval based on the Grey Level Co-occurrence matrices. Second, the automatic fuzzy clustering algorithm for interval data to determine the prior probability for the classification problem by Bayesian method is created. Finally, the new principle to classify for image is established. Combining the above three improvements, we have the effective method to classify the images. In addition, the proposed method can be performed rapidly for the real data by the established Matlab procedure. Four image data sets with the different characters are used to illustrate the proposed method, and to compare to the well-known algorithms like Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Fisher method, Naive Bayes, Multi-Supported Vector Machine (Multi-SVM), Convolutional Neural Networks (CNN), and VGG-19. The results show that the proposed method has the good and stable empirical error, and give the outstanding result about time cost.

Social media platform such as WeChat provides rich cover images for covert communication by steganography. However, in order to save band-width, storage space and make images load faster, the images often will be compressed, which makes the image steganography algorithms designed for lossless network channels unusable. Based on DCT and SVD in nonsubsampled shearlet transform domain, a robust JPEG steganography algorithm is proposed, which can resist image compression and correctly extract the embedded secret message from the compressed stego image. First, by combining the advantages of nonsubsampled shearlet transform, DCT and SVD, the construction method for robust embedding domain is proposed. Then, based on minimal distortion principle, the framework of the proposed robust JPEG steganography algorithm is given and the key steps are described in details. The experimental results show that the proposed JPEG steganography algorithm can achieve competitive robustness and anti-detection capability in contrast to the state-of-the-art robust steganography algorithms. Moreover, it can extract the secret message correctly even if the stego image is compressed by WeChat.

The product quality is the major factor for enhancing the production ability and competitiveness. Decreasing the cost and increasing production capacity are common approaches to realize the enhancement of the product quality. The production managers apply various multimedia data to evaluate the product quality. For example, capturing the stamping sound to evaluate the correct cutting and taking the component image to measure the chip positions are common heterogeneous multimedia data that are applied to manufacturing. However, the production managers prefer to minimize the number of defective products, e. g. the secondary operation and fixing the product tolerance in the assembly stage, to fitting the production target. Therefore, contrasting the defective product identification procedure with high accuracy becomes a challenge due to the decrease of the number of the defective products. In this paper, we propose the Rule Classification with Oversampling (RCOS) approach to provide the high accuracy with few defective products. The proposed RCOS includes the oversampling technique and the rule classification approach to emphasize the properties of the defective products and provide the precise classes. Given few defective products, capturing the properties of the failure is difficult. The RCOS considers the revised Synthetic Minority Over-Sampling Technique (SMOTE) to highlight the failure properties, and then the rule model is considered to extract the root cause of the defective products. We implement the proposed RCOS in the semiconductor production line. From the experiment results, the proposed RCOS provide about at most 98% in accuracy, and the comparison shows that the results have been improved in common criteria e. g. the true-positive rate, G mean, F1 score, and False Alarm Rate. Therefore, the proposed RCOS provides high practicality for the implementation consideration.

Although particle swarm optimization (PSO) algorithm shows excellent performance in solving optimization problems, how to balance exploration and exploitation is still a crucial problem. In this paper, an adaptive comprehensive learning particle swarm optimization with spatial weighting (APSO-SW) is proposed. In APSO-SW, in order to increase the population diversity, the Euclidean distance between each particle and the global optimum is calculated in each generation and the particles in the whole population select their exemplars learning weight according to their Euclidean distance adaptively. Therefore, not only different particles have various learning weight, but also the same particle can select learning weight adaptively in different generations. In addition, for the purpose of improving the convergence property, the terminal elimination strategy is used. In terminal elimination strategy, the population can delete inferior particles and add preferable particles dynamically during the process of evolution. The comparisons among APSO-SW and other 7 state-of-the-art PSO variants on the CEC2013 and CEC2017 test suites reflect that APSO-SW is a trustable and remarkable optimization algorithm for solving various types problems. Furthermore, extensive experiments confirm the validity of our method.

Existing image databases contain a few diversity of images. Likewise, there is no specific image base available in other situations, leading to the need to undertake additional efforts in capturing images and creating datasets. Many of these datasets contain only a single object in each image, but often the scenario in which projects must operate in production requires several objects per image. Thus, it is necessary to expand original datasets into more complex ones with specific combinations to achieve the goal of the application. This work proposes a technique for image generation to extend an initial dataset. It has been designed generically to work with various images and create a data set from some initial images. The generated set of images is used in a distributed environment. It is possible to perform image generation in this environment, producing datasets with specific images to work in certain applications. The generation of images consists of two methods: generation by deformation and generation by a neural network. With the proposed methods, this work sought to bring as main contributions the specification and implementation of an image generating component so that it is possible to easily integrate it with possible heterogeneous devices capable of parallel computing, such as General Purpose Graphics Processing Unit (GPGPU). In comparison with the existing methods to the proposed one, this one proposes to use the image generator enlarging an initial image bank with the combination of two methods. Some experiments are presented doing generation with handwritten digits to validate the proposed approach. The generator was designed with CUDA and GPU-optimized libraries as TensorFlow-specific modules. The results obtained can optimize the integration process with the simulation of possible stimuli choices, avoiding problems in the generation of image phase tests.

Walking is the most common physical activity in humans, and gait can be used as a measure of human health. If the gait is unnatural or uncomfortable, it indicates a problem inside or outside the person’s body. In particular, for the elderly, walking is used as an important indicator of their health status. In this study, we developed an algorithm that can determine whether human walking is natural or unnatural, by comparing the autocorrelation coefficients of the left and right foot. We used F1-scores to measure the accuracy of the gait result determined by the algorithm. Natural walking was accurately distinguished with 80% accuracy, and unnatural with 60% accuracy. Owing to the splint attached to one foot to express unnatural walking, both feet affected gait, resulting in slightly lower accuracy than natural walking. As a future study, it is possible to devise a method to improve accuracy by extracting various gait features that can be obtained through gait and using artificial intelligence algorithms such as machine learning or deep learning.

Recently, human facial attributes analysis has become an important research topic in the field of pattern recognition and computer vision. In fact, various tasks reveal related but different patterns between facial age attribute, race attribute, and gender attribute. Therefore, it is important to construct a facial multi-attribute estimation model to reveal the relationship between different attributes. However, on the one hand, there are some drawbacks in existing facial datasets, such as the lack of some attribute labels or incomplete attribute distribution, so it is infeasible to realize facial multi-attribute estimation on single facial dataset at the same time. On the other hand, in different datasets facial attributes features and labels tend to be heterogeneous, the distribution divergence and the dimension differences due to the changes in collection equipment and image resolution. To this end, this work first proposes the Cross-dataset heterogeneous Adaptation learning facial multiple attributeS joint Estimation (CASE) to mitigate distribution divergence among different facial attributes. Firstly, this work adopts different coding strategies for different face attributes, to maintain the inherent attributes of face attributes. Secondly, in order to explore the potential relationship between labels of different attributes, labels of different attributes are merged and the output relation regularization term for multi-label mapping projection is constructed. Finally, extensive experiments have testified the effectiveness and superiority of the proposed methods.