Search

In the dim-small target detection field, background suppression is a key technique for stably extracting the target. In order to effectively suppress the background to enhance the target, this paper presents a novel background modeling algorithm, which constructs base functions for each pixel based on the local region background and models the background of each pixel, named single pixel background modeling (SPB). In SPB, the low-rank blocks of the local backgrounds are first obtained to construct the background base functions of the center pixel. Then, the background of the center pixel is optimally estimated by the background bases. Experiments demonstrate that in the case of extremely low signal-to-noise ratio (SNR < 1.5 dB) and complex motion state of targets, SPB can stably and effectively separate the target from the strongly undulant sky background. The difference image obtained via SPB background modeling has the characters: the non-target residual could be white noise, and the target is significantly enhanced. Compared with the other typical five algorithms, SPB remarkably outperforms other algorithms to detect the target of a low signal-to-noise ratio.

In this paper, a mutually enhanced modeling method (MEMe) is presented for human pose estimation, which focuses on enhancing lightweight model performance, but with low complexity. To obtain higher accuracy, a traditional model scale is largely expanded with heavy deployment difficulties. However, for a more lightweight model, there is a large performance gap compared to the former; thus, an urgent need for a way to fill it. Therefore, we propose a MEMe to reconstruct a lightweight baseline model, EffBase transferred intuitively from EfficientDet, into the efficient and effective pose (EEffPose) net, which contains three mutually enhanced modules: the Enhanced EffNet (EEffNet) backbone, the total fusion neck (TFNeck), and the final attention head (FAHead). Extensive experiments on COCO and MPII benchmarks show that our MEMe-based models reach state-of-the-art performances, with limited parameters. Specifically, in the same conditions, our EEffPose-P0 with 256 × 192 can use only 8.98 M parameters to achieve 75.4 AP on the COCO val set, which outperforms HRNet-W48, but with only 14% of its parameters.

Neural network pruning, an important method to reduce the computational complexity of deep models, can be well applied to devices with limited resources. However, most current methods focus on some kind of information about the filter itself to prune the network, rarely exploring the relationship between the feature maps and the filters. In this paper, two novel pruning methods are proposed. First, a new pruning method is proposed, which reflects the importance of filters by exploring the information in the feature maps. Based on the premise that the more information there is, more important the feature map is, the information entropy of feature maps is used to measure information, which is used to evaluate the importance of each filter in the current layer. Further, normalization is used to realize cross layer comparison. As a result, based on the method mentioned above, the network structure is efficiently pruned while its performance is well reserved. Second, we proposed a parallel pruning method using the combination of our pruning method above and slimming pruning method which has better results in terms of computational cost. Our methods perform better in terms of accuracy, parameters, and FLOPs compared to most advanced methods. On ImageNet, it is achieved 72.02% top1 accuracy for ResNet50 with merely 11.41M parameters and 1.12B FLOPs.For DenseNet40, it is obtained 94.04% accuracy with only 0.38M parameters and 110.72M FLOPs on CIFAR10, and our parallel pruning method makes the parameters and FLOPs are just 0.37M and 100.12M, respectively, with little loss of accuracy.

Blind image deblurring, also known as blind image deconvolution, is a long-standing challenge in the field of image processing and low-level vision. To restore a clear version of a severely degraded image, this paper proposes a blind deblurring algorithm based on the sigmoid function, which constructs novel blind deblurring estimators for both the original image and the degradation process by exploring the excellent property of sigmoid function and considering image derivative constraints. Owing to these symmetric and non-linear estimators of low computation complexity, high-quality images can be obtained by the algorithm. The algorithm is also extended to image sequences. The sigmoid function enables the proposed algorithm to achieve state-of-the-art performance in various scenarios, including natural, text, face, and low-illumination images. Furthermore, the method can be extended naturally to non-uniform deblurring. Quantitative and qualitative experimental evaluations indicate that the algorithm can remove the blur effect and improve the image quality of actual and simulated images. Finally, the use of sigmoid function provides a new approach to algorithm performance optimization in the field of image restoration.

Obtaining accurate global motion is a crucial step for video stabilization. This paper proposes a robust and simple method to implement global motion estimation. We don’t extend the framework of 2D video stabilization but add a “plug and play” module to motion estimation based on feature points. Firstly, simple linear iterative clustering (SLIC) pre-segmentation is used to obtain superpixels of the video frame, clustering is performed according to the superpixel centroid motion vector and cluster center with large value is eliminated. Secondly, in order to obtain accurate global motion estimation, an improved K-means clustering is proposed. We match the feature points of the remaining superpixels between two adjacent frames, establish a feature points’ motion vector space, and use improved K-means clustering for clustering. Finally, the richest cluster is being retained, and the global motion is obtained by homography transformation. Our proposed method has been verified on different types of videos and has efficient performance than traditional approaches. The stabilization video has an average improvement of 0.24 in the structural similarity index than the original video and 0.1 higher than the traditional method.

Most of the existing trackers address the visual tracking problem by extracting an appearance template from the first frame, which is used to localize the target in the current frame. Unfortunately, they typically face the model degeneration challenge, which easily results in model drift and target loss. To address this issue, a novel Template Adjustment Siamese Network (TA-Siam) is proposed in this paper. The proposed framework TA-Siam consists of two simple subnetworks: The template adjustment subnetwork for feature extraction and the classification-regression subnetwork for bounding box prediction. The template adjustment module adaptively uses the feature of subsequent frames to adjust the current template. It makes the template adapt to the target appearance variation of long-term sequence and effectively overcomes model drift problem of Siamese networks. In order to reduce classification errors, the rhombus labels are proposed in our TA-Siam. For more efficient learning and faster convergence, our proposed tracker uses a more effective regression loss in the training process. Extensive experiments and comparisons with trackers are conducted on the challenging benchmarks including VOT2016, VOT2018, OTB50, OTB100, GOT-10K, and LaSOT. Our TA-Siam achieves state-of-the-art performance at the speed of 45 FPS.

Graph convolutional networks (GCNs) have brought considerable improvement to the skeleton-based action recognition task. Existing GCN-based methods usually use the fixed spatial graph size among all the layers. It severely affects the model’s abilities to exploit the global and semantic discriminative information due to the limits of receptive fields. Furthermore, the fixed graph size would cause many redundancies in the representation of actions, which is inefficient for the model. The redundancies could also hinder the model from focusing on beneficial features. To address those issues, we proposed a plug-and-play channel adaptive merging module (CAMM) specific for the human skeleton graph, which can merge the vertices from the same part of the skeleton graph adaptively and efficiently. The merge weights are different across the channels, so every channel has its flexibility to integrate the joints. Then, we build a novel shallow graph convolutional network (SGCN) based on the module, which achieves state-of-the-art performance with less computational cost. Experimental results on NTU-RGB+D and Kinetics-Skeleton illustrates the superiority of our methods.

Microwave Staring Correlated Imaging (MSCI) is a newly proposed computational high-resolution imaging technique. The imaging performance of MSCI with the existence of modeling errors depends on the properties of the imaging matrix and the relative perturbation error resulted from existing errors. In conventional transient-radiation-fields-based MSCI, which is commonly accomplished by utilizing random frequency-hopping (FH) waveforms, the multiple transmitters should be controlled individually and simultaneously. System complexity and control difficulty are hence increased, and various types of modeling errors are introduced as well. The computation accuracy of radiation fields is heavily worsened by the modeling errors, and the transient effect makes it hard to take direct and high-precision measurements of the radiation fields and calibrate the modeling errors with the measuring result. To simplify the system complexity and reduce error sources, in this paper, steady-radiation-fields-sequence-based MSCI (SRFS-MSCI) method is proposed. The multiple transmitters are excited with coherent signals at the same observation moment, with the signal frequency varying in the whole frequency band during the imaging process. By elaborately designing the array configuration and the amplitude and phase sequences of the coherent transmitters, the SRFS-MSCI is thus implemented. Comparing the system architecture of the proposed SRFS-MSCI with the conventional random FH-based MSCI, it can be found that the proposed method significantly reduces the number of baseband modules and simplifies the system architecture and control logic, which contributes to reducing error sources such as baseband synchronization errors and decreasing deterioration caused by error cascade. To further optimize the design parameters in the proposed SRFS-MSCI system, the Simulated Annealing (SA) algorithm is utilized to optimize the amplitude sequences, the phase sequences, and the antenna positions individually and jointly. Numerical imaging experiments and real-world imaging experiment demonstrate the effectiveness of the proposed SRFS-MSCI method that recognizable high-resolution recovery results are obtained with simplified system structure and optimized system parameters.

SiamFC has a simple network structure and can be pretrained offline on a large data set, so it has attracted the attention of many researchers. It has no online learning process at all. Hence, there are no good solutions for some complex tracking scenarios such as occlusion and large target deformation. For this problem, we propose a method using the Kalman filter method and fusion multiresolution features and get multiple response scores. The Kalman filter acquires the target’s trajectory information, which is used to process complex tracking scenes and to change the selection method of the search area. This also enables our tracker to stably track fast moving targets.The introduction of the Kalman filter compensates for the shortcomings that SiamFC can only track offline, and the tracking network has an online learning process. The fusion of multiresolution features to obtain multiple response scores map helps the tracker to obtain robust features that can be adapted to a variety of tracking targets. Our proposed method has reached the state-of-the-art in testing on five data sets and can be run in real time (40 fps), including OTB2013, OTB2015, OTB50, VOT2015 and VOT 2016.

In the reconstruction of sparse signals in compressed sensing, the reconstruction algorithm is required to reconstruct the sparsest form of signal. In order to minimize the objective function, minimal norm algorithm and greedy pursuit algorithm are most commonly used. The minimum L1 norm algorithm has very high reconstruction accuracy, but this convex optimization algorithm cannot get the sparsest signal like the minimum L0 norm algorithm. However, because the L0 norm method is a non-convex problem, it is difficult to get the global optimal solution and the amount of calculation required is huge. In this paper, a new algorithm is proposed to approximate the smooth L0 norm from the approximate L2 norm. First we set up an approximation function model of the sparse term, then the minimum value of the objective function is solved by the gradient projection, and the weight of the function model of the sparse term in the objective function is adjusted adaptively by the reconstruction error value to reconstruct the sparse signal more accurately. Compared with the pseudo inverse of L2 norm and the L1 norm algorithm, this new algorithm has a lower reconstruction error in one-dimensional sparse signal reconstruction. In simulation experiments of two-dimensional image signal reconstruction, the new algorithm has shorter image reconstruction time and higher image reconstruction accuracy compared with the usually used greedy algorithm and the minimum norm algorithm.