International Journal of Pattern Recognition and Artificial Intelligence

Flame detection is of great significance in a fire prevention system. YOLOv4 has poor real-time performance on flame detection caused by the complex structure and high parameter size. To address this problem, a novel flame detection framework, YOLO for flame (YOLO-F), is proposed in this paper. The backbone of YOLOv4 is simplified from the original 53 convolutional layers to 34 convolutional layers to reduce the number of parameters by simplifying the structure of the CSPBlock. Based on the FPN, an effective and light-weight feature pyramid architecture, namely FPNs-SE, is then proposed and the neck part of YOLOv4 is replaced by FPNs-SE to enhance the feature extraction ability of different scales. In addition, the CIoU loss in the YOLOv4 ignores the similarity measure of the area between the predicted bounding box and the ground-truth bounding box. An effective loss named ACIoU is proposed in this paper in order to handle the above issue and further improve the detection accuracy. The proposed methods are tested on FLAME dataset and network crawled dataset, respectively. The mAP, recall, and precision of YOLO-F are higher by 2.01%, 4.0%, 2.0% on average than those of YOLOv4. With input size of $416\times 416$ and on a single GTX 1660, the operating speed of our method can reach 24.53fps, which is improved by 38.04% compared with YOLOv4. The experimental results show that our method is more robust to the small flame and flame-like objects and can achieve the best balance of detection speed and accuracy. The code is made available at https://github.com/Windxy/YOLO-F.

Human pose estimation (HPE) is a research hotspot in the field of computer vision. Most of the existing approaches first generate low-resolution representation from high-resolution representation through continuous serial downsampling, and then reconstruct high-resolution results from low-resolution features through continuous serial upsampling, which loses a lot of effective feature information and leads to slow model inference. In this paper, the Fast Accuracy Network (FANet), a framework that enables fast and high-accuracy HPE, is proposed. The innovation lies in that, first of all, a grid structure is proposed and adopted, which can be regarded as a set of deep paths and shallow paths. The structure uses multiple high-resolution and low-resolution branch pairs to perform skip-level connections at different scale-space levels so that the information can be exchanged between different resolution representations for many times. The feature information fusion of multi-scale space is realized to obtain more abundant feature information. Second, an improved bottleneck block is proposed to extract effective feature information with fewer parameters, ensuring that the computational burden is reduced without sacrificing accuracy performance. The experimental results show that, compared with other current models, FANet has faster inference speed on the premise of a slight improvement in accuracy performance.

We propose an improved YOLOV4-tiny traffic sign recognition algorithm for easy deployment on mobile or embedded devices to address the problems of a large number of parameters, low recognition accuracy in complex scenarios. The model uses the YOLOV4-tiny network as the basic framework. First, Octave Convolution is introduced into the backbone network to reduce the redundancy of low-frequency features. Second, the convolutional block attention module is used to strengthen the weights of traffic sign regions and reduce the weights of invalid features. Finally, the Feature Pyramid Network structure is replaced by the Simplified Path Aggregation Network structure in the feature fusion stage to enhance the feature information and further reduce the miss detection rate. The experiment proves that our method outperforms YOLOV4-tiny in recognition accuracy and detection speed on the TT100K dataset, and can easily meet the requirements of traffic sign recognition.

Learning from only few samples is a challenging problem and meta-learning is an effective approach to solve it. Meta-learning model aims to learn by training a large number of other samples. When encountering target task, the model can quickly adapt and obtain better performance with only few labeled samples. However, general meta-learning only provides a universal model that has certain generalization ability for all unknown tasks, which causes limited effects on specific target tasks. In this paper, we propose a Few-shot Directed Meta-learning (FSDML) model to specialize and solve the target task by using few labeled samples of the target task to direct the meta-learning process. FSDML divides model parameters into shared parameters and target adaptation parameters to store prior knowledge and determine the update direction. These two parts of the parameters are updated in different stages of training. We conduct experiments of image classification task on miniImageNet and Omniglot and the results show that FSDML has better performance.

Aiming at the “bottleneck” problems of the traditional network security situation awareness model, such as large equipment limitations, single data source and poor integration ability, weak level of autonomous learning and data mining, a network security situation awareness framework suitable for big data is constructed. A gate recurrent unit (GRU) model is established to effectively extract features from the situation data set through the deep learning algorithm of big data. It is a method to automatically mine and analyze the hidden relationship and change trend of network security situation, realize the high-speed acquisition and fusion of massive multi-source heterogeneous data, and perceive the network security situation from an all-round perspective. The experimental results show that this method has a good awareness effect on network threats, and has strong representation ability in the face of network threats. It can effectively perceive the network threat situation without relying on data labels, which verifies that this method can effectively improve the efficiency and accuracy of security situation awareness.

Synthetic Aperture Radar (SAR) represents a type of active remote sensing technology that uses microwave electromagnetic radiation to produce and send data to the surface of a target location. SAR imaging is frequently used in national security applications since it is unaffected by weather, geographical location, or time. In this system, many approaches are examined, to improve automation for segmentation and classification. The utilization of Deep Neural Networks (DNNs) to classify SAR images has gotten a lot of attention, and it usually requires several layers of deep models for feature learning. With insufficient training data, however, the DNN will get affected by the overfitting issue. The major purpose of this work is to make a development on introducing a new framework for SAR image segmentation and categorization using deep learning. Owing to the coherent nature of the backscattering signal, SARs create speckle noise in their images. If the image has noisy material, classification becomes more challenging. Hence, the pre-processing of the images is employed by linear spatial filtering to remove the noise. Further, the Optimized U-Net is used for the segmentation. For the segmented images, the Binary Robust Independent Elementary Features (BRIEF) concept is adopted as the feature descriptor. These features are inputted to the Convolutional Neural Network (CNN) with Tuned Weight DNN (C-TWDNN) for the classification. In both segmentation and classification, the parameter tuning is employed by the combination of Galactic Swarm Optimization (GSO) and Deer Hunting Optimization Algorithm (DHOA) called the Self-adaptive-Galactic Deer Hunting Optimization (SA-GDHO). Experiments are conducted on a variety of public datasets, demonstrating that our method is capable of outperforming various expert systems and deep structured architectures.

Fingerprints being the most widely employed biometric trait, due to their high acceptability and low sensing cost, have replaced the traditional methods of human authentication. Although, the deployment of these biometrics-based recognition systems is accelerating, they are still susceptible to spoofing attacks where an attacker presents a fake artifact generated from silicone, candle wax, gelatin, etc. To safeguard sensor modules from these attacks, there is a requirement of an anti-deception mechanism known as fingerprint spoof detectors (FSD) also known as anti-spoofing mechanisms. A lot of research work has been carried out to design fingerprint anti-spoofing techniques in the past decades and currently, it is oriented towards deep learning (DL)-based modeling. In the field of fingerprint anti-spoofing, since the 2014, the paradigm has shifted from manually crafted features to deep features engineering. Hence, in this study, we present a detailed analysis of the recent developments in DL based FSDs. Additionally, we provide a brief comparative study of standard evaluation protocols that include benchmark anti-spoofing datasets as well as performance evaluation metrics. Although significant progress has been witnessed in the field of DL-based FSDs, still challenges are manifold. Therefore, we investigated these techniques critically to list open research issues along with their viable remedies that may put forward a future direction for the research community. The majority of the research work reveals that deep feature extraction for fingerprint liveness detection demonstrates promising performance in the case of cross-sensor scenarios. Though convolution neural network (CNN) models extract deep-level features to improve the classification accuracy, their increased complexity and training overhead is a tradeoff between both the parameters. Furthermore, enhancing the performance of presentation attack detection (PAD) techniques in the cross-material scenario is still an open challenge for researchers.

In recent years, with the development of remote sensing technology and the enhancement of the value of remote sensing images in military and civil fields, remote sensing image object segmentation has also received more and more attention. This paper mainly studies the application of instance segmentation based on deep convolutional neural network in the remote sensing image. This paper proposes an attention balanced feature pyramid module, which strengthens multi-level features and uses the attention module to suppress the interference features of noise in the complex background. In addiction, Soft-NMS is introduced to improve the performance of the network, and GIoU loss is introduced to improve the effect of object detection. The proposed network improves the average detection and segmentation accuracy (mAP) values from $41.75\%$ and $35.34\%$ to $43.05\%$ and $36.02\%$, respectively.

Video inpainting aspires to fill the Spatio-temporal holes in videos with probable and coherent content. This process recovers the missing content of corrupted video effectively, which is useful in many fields, including removal of watermarking and video restoration. The difficulties of creating video contents with exquisite detail while maintaining spatiotemporal coherence in the missing areas is the main difficulty in the video inpainting process. Modern studies ignore semantic structural coherence maintenance between frames in favor of using flow information to synthesize temporally smooth pixels. In this paper, Political Improved Invasive Weed Optimization (PIIWO)-based optimal exemplar is designed for the productive video inpainting process. Accordingly, the developed PIIWO algorithm is newly designed by combining Political Optimizer (PO) and Improved Invasive Weed Optimization (IIWO). Here, the inpainting results obtained from context-aware Ant Lion Gray Wolf Optimization (ALGWO)-based Markov Random Field (MRF) modeling, Whale Monarch Butterfly Optimization (Whale MBO)-based Deep Convolutional Neural Network (DCNN), K-Nearest Neighbors (KNN) with Bhattacharya distance, Bi-harmonic function modules and developed PIIWO-based exemplar model are fused using Bayes probabilistic fusion for producing the final result. Three metrics, peak signal-to-noise ratio (PSNR), second derivative like the measure of enhancement (SDME) and structural similarity (SSIM) of 40.19dB, 78.07dB and 0.9857, respectively, are used to assess the performance of the developed video inpainting technique.

Due to the complexity and diversity of indoor environment objects and interference occlusions, the accuracy of multi-object target detection based on 3D point cloud is limited. To address this issue, we present a multi-target detection method based on adaptive feature adjustment (AFA) of 3D point cloud. First, our method preprocesses the dataset and constructs a backbone module. Afterwards, our method uses an improved PointNet$++$ network for feature adaptive learning, where an AFA module is added to learn the influence relationship between point pairs. The proposed method then establishes the relationship between contexts in the local point set area and extracts the feature of point cloud. Using the idea of Hough voting, our method can generate some votes close to the particle. Using these votes to generate proposal, the proposed method adds CBAM attention mechanisms to both modules of voting and proposal, which can fuse the feature information of the channel and expand the receptive field in space. Our method can enhance the important features and weaken the unimportant features, making the extracted features more directional and enhancing the expressiveness of the network. Finally, the generated results are visualized to complete the multi-target detection of 3D point cloud. To verify the effectiveness of our proposed method, two large datasets with real 3D scanning, scanNet2 and SunRGB-D, are used for training the network. The experimental results show that the proposed method can improve the effectiveness of point cloud target detection in indoor scenes, getting a higher detection accuracy.

Compared to cross-entropy in deep learning, triplet loss is less affected by the biased label information and widely used in fine-grained visual tasks. Especially in person re-identification (re-ID), triplet loss is improved with batch-hard sampling which only selects the hardest samples during the training process to reduce invalid triplets involved in the loss computation. The hardest samples’ loss computation can provide a more intense gradient descent than raw samples. However, the batch-hard triplet loss discards multiple samples with important information, which can negatively impact feature learning. Besides, the hardest samples cause loss stuck problems frequently in training. In this work, we propose a balanced triplet loss for comprehensive feature learning and stable model convergence. The balanced triplet loss only mines the hardest negative samples of each category within a mini-batch. Compared with batch-hard triplet loss, it preserves the features of all the negative categories rather than one negative category with the hardest negative sample. It achieves a balance between triplet selection and information loss. The experiments show that our method can produce competitive results in re-ID tasks. In addition, we analyze the correlation between the intensity of data mining and the granularity of feature learning and further adapt the balanced triplet loss to general fine-grained image classification. The experiments prove the adapted balanced triplet loss also outperforms cross-entropy in multiple datasets of different scales.

The degree of insulation aging of power cables is closely related to their partial discharge (PD) level, so the analysis of PD signals can be used to realize the cable condition detection. However, after performing online detection of PDs on power cables, the collected signals always contain interference signals due to the influence of electromagnetic interference in the field. In order to identify each type of local discharge signal from the interference signal, this paper proposes a clustering identification algorithm for local discharge signals, which mainly involves pulse extraction, feature parameter extraction and clustering identification process. The algorithm first extracts the pulse signal by combining the amplitude–time threshold method and the time domain energy method, then obtains the feature vector of the signal according to the synchronous multi-channel method, designs a fuzzy C-mean clustering algorithm based on subtractive clustering to determine the initial clustering center to cluster the samples and finally analyzes and checks the clustering results according to the phase resolved PD (PRPD) of a single class of signals and the fit of the two-parameter Weibull distribution function. The clustering results were analyzed and examined. The experimental results show that the proposed algorithm can extract pulse signals efficiently and accurately, and the synchronous multi-channel method can characterize pulse signals better. Meanwhile, the algorithm can determine the optimal number of classes adaptively according to the clustering effectiveness function and adopt subtractive clustering to initialize the clustering center, which can approach the optimal solution faster, and can effectively cluster a variety of discharge signals, so as to realize the type identification of single-class discharge signals.

In this paper, we adopt two methods to study the problem. Initially, directly from the second-order differential equation, we obtain a sufficient condition (SC) for the mean square exponential stability (MSES) of the system at the equilibrium point by constructing a suitable function and applying some properties of calculus. Thereafter, the system is transformed into a vector form, using the basic solution matrix of linear differential equation, constructing a piecewise function and using the generalized Halanay one-dimensional delay differential inequality, another SC is given for the P-moment exponential stability (PMES) of the system at the equilibrium point. Finally, two examples are used to investigate the correctness and demonstrate that each SC has own advantage, the suitable theorem can be selected according to the parameters.

The classification is usually degraded due to the imbalanced class distribution. Synthetic minority oversampling technique (SMOTE) has been successful in improving imbalanced classification and has received great praise. Overgeneralization is one of the most challenges in SMOTE. Although multiple SMOTE-based variations are proposed against overgeneralization, they still have the following shortcomings: (a) creating too many synthetic samples in high-density regions; (b) removing suspicious noise directly instead of modifying them; (c) relying on many parameters. This paper proposes a new SMOTE based on adaptive noise optimization and fast search for local sets (SMOTE-ANO-FLS) to overcome the overgeneralization and the shortcomings of existing works. First, SMOTE-ANO-FLS uses the $k$-D tree to fast search the local sets for each sample. Second, a new noise detection method based on local sets and the imbalanced ratio is proposed to detect suspicious noise. Third, a new adaptive noise optimization method is proposed to modify detected suspicious noise instead of removing them. Finally, a new probability weight based on local sets is proposed to help create more synthetic minority class samples in borderline and sparse regions. The effectiveness of SMOTE-ANO-FLS is proven by employing 7 oversampling methods and random forest on the extensive synthetic and real data sets.