Search

The proliferation of remote sensing image data in recent years has posed a pressing need for efficient compression techniques due to constrained transmission bandwidth. While lossless compression preserves image fidelity, it falls short of meeting real-time demands. Conversely, conventional lossy compression methods can attain high compression ratios for real-time applications, but often introduce issues like block artifacts, blurring, and distortions in the decompressed images. Hence, we propose the Global Feature-Assisted Segmentation Compression Network (GFSCompNet) as a solution for high compression ratio lossy compression. Initially, we design a segmentation network utilizing a dual-branch global feature-assisted segmentation approach to precisely detect small targets in remote sensing images. On the compression side of the network, we leverage an attention mechanism and code rate allocation technique to seamlessly merge the segmented small target information with the original image, thereby allocating a higher compression code rate to the small target region. Furthermore, a joint hyper-priority decoding and entropy coding estimation network is proposed to further remove the redundancy in the potential representation and improve the compression ratio. Experimental results conducted under conditions of high compression ratios and comparable bit rates demonstrate that our approach yields higher-quality reconstructed images compared to the JPEG algorithm and outperforms other deep learning-based image compression methods. Additionally, it effectively preserves small target information, thereby enhancing the interpretability of machine learning models. [ABSTRACT FROM AUTHOR]

Violent behavior detection (VioBD), as a special action recognition task, aims to detect violent behaviors in videos, such as mutual fighting and assault. Some progress has been made in the research of violence detection, but the existing methods have poor real-time performance and the algorithm performance is limited by the interference of complex backgrounds and the occlusion of dense crowds. To solve the above problems, we propose an end-to-end real-time violence detection framework based on 2D CNNs. First, we propose a lightweight skeletal image (SI) as the input modality, which can obtain the human body posture information and richer contextual information, and at the same time remove the background interference. As tested, at the same accuracy, the resolution of SI modality is only one-third of that of RGB modality, which greatly improves the real-time performance of model training and inference, and at the same resolution, SI modality has higher inaccuracy. Second, we also design a parallel prediction module (PPM), which can simultaneously obtain the single image detection results and the inter-frame motion information of the video, which can improve the real-time performance of the algorithm compared with the traditional “detect the image first, understand the video later" mode. In addition, we propose an auxiliary parameter generation module (APGM) with both efficiency and accuracy, APGM is a 2D CNNs-based video understanding module for weighting the spatial information of the video features, processing speed can reach 30–40 frames per second, and compared with models such as CNN-LSTM (Iqrar et al., Aamir: Cnn-lstm based smart real-time video surveillance system. In: 2022 14th International Conference on Mathematics, Actuarial, Science, Computer Science and Statistics (MACS), pages 1–5. IEEE, 2022) and Ludl et al. (Cristóbal: Simple yet efficient real-time pose-based action recognition. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 581–588. IEEE, 1999), the propagation effect speed can be increased by an average of 3 ∼ 20 frames per second per group of clips, which further improves the video motion detection efficiency and accuracy, greatly improving real-time performance. We conducted experiments on some challenging benchmarks, and RVBDN can maintain excellent speed and accuracy in long-term interactions, and are able to meet real-time requirements in methods for violence detection and spatio-temporal action detection. Finally, we update our proposed new dataset on violence detection images (violence image dataset). Dataset is available at [ABSTRACT FROM AUTHOR]

Due to the influence of light scattering, absorption and noise, the underwater environment presents a range of challenges for image processing tasks, such as color cast, low contrast and poor readability. It is truly regrettable that most algorithms primarily rely on synthetic datasets or few real-world river underwater images for evaluation. To address these challenges, we build a river underwater image enhancement dataset and propose an enhancement approach that leverages multi-color space, prior knowledge of underwater imaging and two-stage adaptive fusion. First, the shallow preprocessing module utilizes deformable convolution to better extract cross information. Next, the parallel RGB and HSV enhancement channels (with attention mechanism) can effectively highlight image characteristics in different colorspaces. Considering the underwater imaging prior, our method reduces the probability of artifacts. Finally, we merge the different frequency features in the first stage. The HSV and RGB components are fused adaptively in the second stage. Extensive experimental results verify the effectiveness of our algorithm in enhancing real-world underwater images. [ABSTRACT FROM AUTHOR]

Underwater images suffer from color cast and low contrast due to the light absorption and scattering. Especially when natural light is not sufficient, large dark areas appear in the captured image, making it impossible to understand the image content. To address this issue, we propose an underwater low-light enhancement method based on Retinex theory. Our model is an end-to-end trainable. The decomposition network decomposes the raw image into reflectance and illumination according to Retinex theory. In the reflectance enhancement network, cross-residual blocks and dense connections can improve the efficiency of feature utilization and the hybrid attention concentrate on the regions of interest in feature maps from different perspectives. The illumination adjustment network utilizes adaptive frequency convolutional blocks to generate additional band information, which reconstructs the more natural illumination. In order to preserve the color consistency of the enhanced image with the reference image, we project the HSV space into the Cartesian coordinate system and use the Euclidean distance as the color cast loss to constrain the enhancement network. Qualitative and quantitative evaluations on different underwater datasets indicate that our method has the excellent performance and can achieve delightful visual enhancements. [ABSTRACT FROM AUTHOR]

H.265/HEVC (high efficiency video coding) adopts in-loop filters to reduce artifacts such as blocking artifacts, ringing artifacts, but those above-mentioned artifacts are still obvious in the case of large quantization step. This paper presents a novel in-loop filtering mechanism based on three-dimensional (3D) sub-bands in adaptive group of frames and convolutional neural network (CNN) models for further improvement of in-loop filtering capability. Firstly, the video frame sequence is adaptively grouped; then, a group of video frames is decomposed into frequency sub-bands by 3D wavelet transform; afterward, frequency sub-bands with different types are filtered by corresponding CNN model; finally, the group of video frames is synthesized by 3D wavelet inverse transform. We apply the nonlinear mapping ability of CNN into the modification of wavelet coefficients. The sub-bands of high frequency and low frequency are filtered by four offline CNN models with different directional characteristics trained with training data generated by 3D wavelet transform, respectively. The details of the high frequency part are enhanced, while the quality of low-pass image is promoted, so that those artifacts can be effectively alleviated. By means of comparing experiments, the subjective and objective results show that our proposed method has better filtering performance than the in-loop filtering mechanism in HM16.18; especially the video frame quality can be effectively improved in the case of large quantization step. [ABSTRACT FROM AUTHOR]

At present, to realize or improve the quality of experience (QoE) is a major goal for network media transmission service, and QoE evaluation is the basis for adjusting the transmission control mechanism. Therefore, a kind of QoE collaborative evaluation method based on fuzzy clustering heuristic algorithm is proposed in this paper, which is concentrated on service score calculation at the server side. The server side collects network transmission quality of service (QoS) parameter, node location data, and user expectation value from client feedback information. Then it manages the historical data in database through the 'big data' process mode, and predicts user score according to heuristic rules. On this basis, it completes fuzzy clustering analysis, and generates service QoE score and management message, which will be finally fed back to clients. Besides, this paper mainly discussed service evaluation generative rules, heuristic evaluation rules and fuzzy clustering analysis methods, and presents service-based QoE evaluation processes. The simulation experiments have verified the effectiveness of QoE collaborative evaluation method based on fuzzy clustering heuristic rules. [ABSTRACT FROM AUTHOR]

Kernel correlation filter (KCF) algorithm is famous for fast tracking speed and has been used widely, while it is susceptible under some challenging scenes. Aiming at the problem of limited applicable scenes of KCF, an improved target tracking method based on extraction of corner points is proposed. Eight neighborhood template is used to filter corner points, and adaptive threshold approach is introduced to minimize the influence of noise on corner extraction. Moreover, owing to appropriate quantity and strong robustness, adaptive corners can reduce the influence of incomplete extraction on edge points while enhancing extraction speed. Subsequently, histogram of oriented gradient (HOG) and color name (CN) features are extracted, and targets are segmented into blocks to solve the problem that standard KCF is likely to lose targets when they have deformation and scale variation. Under occlusion, learning rate parameter is adjusted adaptively ensuring accuracy of model updating. To get rid of drift phenomenon when targets have fast motion, multiple targets are correlated, and the final center position of each target is achieved according to the contour depicted by generalized Hough algorithm. Experimental results on the dataset and actual scenes demonstrate that our proposed method improves the EAO from 0.299 to 0.505, and improves overall precision of diverse attributes from 0.781 to 0.787. [ABSTRACT FROM AUTHOR]