Progressive Fusion Research Articles

Chondrogenesis mediates progression of ankylosing spondylitis through heterotopic ossification

Ankylosing spondylitis (AS) is chronic inflammatory arthritis with a progressive fusion of axial joints. Anti-inflammatory treatments such as anti-TNF-α antibody therapy suppress inflammation but do not effectively halt the progression of spine fusion in AS patients. Here we report that the autoimmune inflammation of AS generates a microenvironment that promotes chondrogenesis in spine ligaments as the process of spine fusion. Chondrocyte differentiation was observed in the ligaments of patients with early-stage AS, and cartilage formation was followed by calcification. Moreover, a large number of giant osteoclasts were found in the inflammatory environment of ligaments and on bony surfaces of calcified cartilage. Resorption activity by these giant osteoclasts generated marrow with high levels of active TGF-β, which induced new bone formation in the ligaments. Notably, no Osterix+ osteoprogenitors were found in osteoclast resorption areas, indicating uncoupled bone resorption and formation. Even at the late and maturation stages, the uncoupled osteoclast resorption in bony interspinous ligament activates TGF-β to induce the progression of ossification in AS patients. Osteoclast resorption of calcified cartilage-initiated ossification in the progression of AS is a similar pathologic process of acquired heterotopic ossification (HO). Our finding of cartilage formation in the ligaments of AS patients revealed that the pathogenesis of spinal fusion is a process of HO and explained why anti-inflammatory treatments do not slow ankylosing once there is new bone formation in spinal soft tissues. Thus, inhibition of HO formation, such as osteoclast activity, cartilage formation, or TGF-β activity could be a potential therapy for AS.

Analysis of pancreatic segmentation algorithm based on deep learning to improve pancreatic critical region segmentation ability on dual-phase CT

Objective: To investigate the segmentation effects of the deep learning method on CT in the arterial phase and venous phase respectively by using subjective and objective evaluation system, and to investigate the factors that affect the difference between arterial phase and venous phase pancreas segmentation and the related factors affecting the venous pancreas segmentation. Method: A total of 218 cases of pancreatic CT scan data in the Department of Radiology of Peking Union Medical College Hospital from January to November 2019 were retrospectively collected. Each case contained images of arterial and venous phases, and the data were randomly divided into training set (139 cases), validation set (20 cases) and test set (59 cases) according to the ratio of the training and verification set to the test set of 7∶3. The two-stage global local progressive fusion network was trained on the training set, the model parameters of the optimal segmentation effect were found on the validation set, and the test set was predicted and the results were evaluated subjectively and objectively. The Likert 5-point scale was used for subjective evaluation based on the critical regions between pancreas and peripheral organs, while the Dice similarity coefficient (DSC) was used for objective evaluation. The paired t test or Wilcoxon paired rank test was used to compare the differences of subjective and objective scores of the arterial phase and venous phase. Results: For the critical regions of the pancreas at the duodenum, duodenal jejunal flexure, left adrenal gland, portal vein, superior mesenteric vein, splenic artery and splenic vein, the median number of subjective scores in arterial phase were 4(4, 5), 5(4, 5), 5(4, 5), 4(4, 5), 5(4, 5), 5(5, 5) and 4(3, 5)points respectively, the median number(first quartile, third quartile) of subjective scores in venous phase were 4(4, 4), 5(4, 5), 5(4, 5), 5(4, 5), 5(5, 5), 4(3, 4) and 5(5, 5) points respectively,there were statistically significant differences of the median number(first quartile, third quartile) of the subjective scores between the arterial and venous phase for the critical regions of the pancreas at the organs described above (all P<0.05). DSC in the venous phase was slightly higher than that in the arterial phase and the difference was not statistically significant (DSC: 0.932 vs 0.921, P=0.952). Subjective scores in venous phase of the pancreas and duodenal jejunum, stomach, and left adrenal gland with fat gaps were 4.64,4.68 and 4.63 points respectively, and those of the group without fat gaps were 4.56,4.62 and 4.56 points respectively, there were statistically significant differences of the subjective scores in venous phase of the groups with fat gaps or not between the pancreas and the organs described above (t=2.147, 2.112, 2.277, all P<0.05). Except the spleen, the density differences between the critical regions of the pancreas and other surrounding organs were statistically significant in arterial phase and venous phase segmentation (all P<0.05). Conclusion: Dual-phase CT was used to construct a deep learning automatic pancreas segmentation model, and the segmentation effect was evaluated subjectively and objectively. Subjective evaluation was helpful to improve the ability to segment the critical regions of the pancreas in the future.

2D progressive fusion module for action recognition

Network convergence as well as recognition accuracy are essential issues when applying Convolutional Neural Networks (CNN) to human action recognition. Most deep learning methods neglect model convergence when striving to improve the abstraction capability, thus degrading the performances sharply when computing resources are limited. To mitigate this problem, we propose a structure named 2D Progressive Fusion (2DPF) Module which is inserted after the 2D backbone CNN layers. 2DPF fuses features through a novel 2D convolution on the spatial and temporal dimensions called variation attenuating convolution and applies fusion techniques to improve the recognition accuracy and the convergency. Our experiments performed on several benchmarks (e.g., Something-Something V1&V2, Kinetics400 & 600, AViD, UCF101) demonstrate the effectiveness of the proposed method.ARTICLE INFO.

Video super‐resolution with non‐local alignment network

Video super-resolution (VSR) aims at recovering high-resolution frames from their low-resolution counterparts. Over the past few years, deep neural networks have dominated the video super-resolution task because of its strong non-linear representational ability. To exploit temporal correlations, most deep neural networks have to face two challenges: (1) how to align consecutive frames containing motions, occlusions and blurring, and establish accurate temporal correspondences, (2) how to effectively fuse aligned frames and balance their contributions. In this work, a novel video super-resolution network, named NLVSR, is proposed to solve above problems in an efficient and effective manner. For alignment, a temporal-spatial non-local operation is employed to align each frame to the reference frame. Compared with existing alignment approaches, the proposed temporal-spatial non-local operation is able to integrate the global information of each frame by a weighted sum, leading to a better performance in alignment. For fusion, an attention-based progressive fusion framework was designed to integrate aligned frames gradually. To penalize the points with low-quality in aligned features, an attention mechanism was employed for a robust reconstruction. Experimental results demonstrate the superiority of the proposed network in terms of quantitative and qualitative evaluation, and surpasses other state-of-the-art methods by 0.33 dB at least.

A Relational-Database Methodology for Incorporating Consciousness into AI Agents

Consciousness is key to Artificial General Intelligence. It is currently unclear what is consciousness and we expect that by the time it is understood, there will probably be millions of man-years invested in the development of AI agents. This paper addresses the following question. How should AI agents be engineered from now on, such that when it becomes clear how consciousness and subjective experiences are generated, the effort of endowing AI agents and robots existing at that time with consciousness is minimized. As far as we know this question has not been studied heretofore. The paper proposes a methodology and architecture that are based on evolving Probabilistic Relational Databases, and on stepwise progressive data fusion. The methodology and architecture are independent of what consciousness turns out to be, provided that it can be incorporated into AI agents.

Towards Integration of Domain Knowledge-Guided Feature Engineering and Deep Feature Learning in Surface Electromyography-Based Hand Movement Recognition.

As a machine-learning-driven decision-making problem, the surface electromyography (sEMG)-based hand movement recognition is one of the key issues in robust control of noninvasive neural interfaces such as myoelectric prosthesis and rehabilitation robot. Despite the recent success in sEMG-based hand movement recognition using end-to-end deep feature learning technologies based on deep learning models, the performance of today's sEMG-based hand movement recognition system is still limited by the noisy, random, and nonstationary nature of sEMG signals and researchers have come up with a number of methods that improve sEMG-based hand movement via feature engineering. Aiming at achieving higher sEMG-based hand movement recognition accuracies while enabling a trade-off between performance and computational complexity, this study proposed a progressive fusion network (PFNet) framework, which improves sEMG-based hand movement recognition via integration of domain knowledge-guided feature engineering and deep feature learning. In particular, it learns high-level feature representations from raw sEMG signals and engineered time-frequency domain features via a feature learning network and a domain knowledge network, respectively, and then employs a 3-stage progressive fusion strategy to progressively fuse the two networks together and obtain the final decisions. Extensive experiments were conducted on five sEMG datasets to evaluate our proposed PFNet, and the experimental results showed that the proposed PFNet could achieve the average hand movement recognition accuracies of 87.8%, 85.4%, 68.3%, 71.7%, and 90.3% on the five datasets, respectively, which outperformed those achieved by the state of the arts.

Bundle Branch Block Pattern Resolution in Non-Selective His Bundle Pacing: Effect of Programmed AV Delays

We have reported that non-selective but not selective His-bundle pacing (HBP) resolves bundle branch block (BBB) pattern. Characterized by ventricular pre-excitation, the mechanism by which nonselective-HBP resolves BBB is unknown. As A-V delays may determine both duration and timing of ventricular pre-excitation, it was hypothesized that an analysis of P-wave to HBP interval in relation to P-wave to His-bundle electrogram (HBE), may yield further insights into resolution of BBB pattern in NS-HBP. Methods: In 27 patients with RBBB+LAD (n=11) and LBBB (n=16), QRS parameters at AV delays resulting in P-HBP interval P-HBE. Results: At P-33 HBP< P-HBE, the pre-excitation duration was longest (63±13ms), and a more normal pre-excited complex replaced the RBBB, LAD, or LBBB pattern. In addition, the QRS-voltage in L1 was restored (LAD and LBBB), QRS-axis normalized (LAD), ventricular activation times decreased (LBBB) and rightward delay (S wave in L1) resolved (RBBB) (p<0.01). At longer A-V delays with P-HBP interval prolonging incrementally beyond P-HBE interval, underlying BBB emerged progressively, and correlated with both delayed pre-excitation and decrease in pre-excitation duration. Conclusion: The resolution of BBB pattern is associated with maximal pre-excitation in NS-HBP. Delayed ventricular pre-excitation and decreased pre-42 excitation duration, caused by with longer AV delays, result in progressive fusion complexes with increasing BBB pattern. Thus, our data is most consistent with a bundle branch pattern being replaced by a more normal pre-excited QRS pattern, and less consistent with correction of BBB itself.

Progressive fusion of feature sets for optimal classification of MI signal

Brain-computer interface (BCI) is the most popular research topic to the researchers of neuroprosthetics. The ultimate goal of this research is to develop a communication channel between human brains and external devices. Feature extraction is one of the most crucial steps in this research. Combination of different features may improve the classification performance, but in most of the cases straight forward combinations of different features lead to a very poor result. So, it is necessary to combine the orthogonal features and omit the redundant ones. It is a complex and time-consuming process. We have developed two new algorithms to find optimum sets of features for fusion to obtain best possible classification accuracy for both subject-specific (SS) and subject-independent (SI) cases. Experimental results indicate that our proposed algorithms, in general, improve the classification results irrespective of the different methodological setup of BCI processes such as number of input channels and spatial filter.

Progressive fusion of feature sets for optimal classification of MI signal

Progressive fusion of feature sets for optimal classification of MI signal

Cross-Layer Feature Pyramid Network for Salient Object Detection.

Feature pyramid network (FPN) based models, which fuse the semantics and salient details in a progressive manner, have been proven highly effective in salient object detection. However, it is observed that these models often generate saliency maps with incomplete object structures or unclear object boundaries, due to the indirect information propagation among distant layers that makes such fusion structure less effective. In this work, we propose a novel Cross-layer Feature Pyramid Network (CFPN), in which direct cross-layer communication is enabled to improve the progressive fusion in salient object detection. Specifically, the proposed network first aggregates multi-scale features from different layers into feature maps that have access to both the high- and low- level information. Then, it distributes the aggregated features to all the involved layers to gain access to richer context. In this way, the distributed features per layer own both semantics and salient details from all other layers simultaneously, and suffer reduced loss of important information during the progressive feature fusion. At last, CFPN fuses the distributed features of each layer stage-by-stage. This way, the high-level features that contain context useful for locating complete objects are preserved until the final output layer, and the low-level features that contain spatial structure details are embedded into each layer to preserve spatial structural details. Extensive experimental results over six widely used salient object detection benchmarks and with three popular backbones clearly demonstrate that CFPN can accurately locate fairly complete salient regions and effectively segment the object boundaries.

A spatial-channel progressive fusion ResNet for remote sensing classification

In recent years, the panchromatic (PAN) and the multispectral (MS) remote sensing images classification has become a research hotspot. In this paper, we propose a spatial-channel progressive fusion residual network (SCPF-ResNet) for multi-resolution remote sensing classification. Firstly, for the inputs of the proposed network, the interactive data fusion strategy (IDFS) combines generalized-intensity-hue-saturation (GIHS), and discrete wavelet transform (DWT) to interfuse patch pairs of the PAN and the MS images, so as to increase the similarity between them, thus reduce the difference in information between them. Secondly, for the branches of feature extraction, we design an adaptive spatial attention module (ASA-Module) and an adaptive channel attention module (ACA-Module) to strengthen spatial features from both larger-sized with smaller-sized targets and spectral features among channels. Finally, we insert the ASA-Module and ACA-Module into the residual modules to form a triple-branch network and use the common spatial-channel features extracted by the Fusion_Branch to gradually enhance the pure independent features extracted by the PAN_Branch and the MS_Branch, respectively. The experimental results indicate that SCPF-ResNet can achieve competitive performance.

A Progressive Fusion Generative Adversarial Network for Realistic and Consistent Video Super-Resolution.

How to effectively fuse temporal information from consecutive frames remains to be a non-trivial problem in video super-resolution (SR), since most existing fusion strategies (direct fusion, slow fusion, or 3D convolution) either fail to make full use of temporal information or cost too much calculation. To this end, we propose a novel progressive fusion network for video SR, in which frames are processed in a way of progressive separation and fusion for the thorough utilization of spatio-temporal information. We particularly incorporate multi-scale structure and hybrid convolutions into the network to capture a wide range of dependencies. We further propose a non-local operation to extract long-range spatio-temporal correlations directly, taking place of traditional motion estimation and motion compensation (ME&MC). This design relieves the complicated ME&MC algorithms, but enjoys better performance than various ME&MC schemes. Finally, we improve generative adversarial training for video SR to avoid temporal artifacts such as flickering and ghosting. In particular, we propose a frame variation loss with a single-sequence training method to generate more realistic and temporally consistent videos. Extensive experiments on public datasets show the superiority of our method over state-of-the-art methods in terms of performance and complexity. Our code is available at https://github.com/psychopa4/MSHPFNL.

Optimization of the Progressive Image Mosaicing Algorithm in Fine Art Image Fusion for Virtual Reality

The fade-in and fade-out algorithm based on the Bernstein polynomial has certain limitations in image fusion. Therefore, this article proposes a new image fusion algorithm. First, the SIFT algorithm is used to register the images. Second, for the disjointed case of overlapping regions, a progressive image mosaic fusion algorithm in the form of a sine function is proposed. Finally, in order to make the progressive image mosaic fusion algorithm suitable for a variety of overlapping regions, this paper adds segmentation technology. The simulation experiment results show that the algorithm proposed in this paper is in good agreement with the spatial details and texture details of a high-resolution panchromatic image, and the time is shorter, which meets the real-time requirements. In addition, the algorithm proposed in this paper is effective in applications such as virtual reality and art image fusion.

Temperament and Character Profiles of Ankylosing Spondylitis Patients Compared With Major Depression Patients and Healthy Controls.

Depressive disorder is common in patients with ankylosing spondylitis (AS), who exhibit specific brain imaging abnormalities resembling those in major depression (MD). We compared temperament and character profiles between patients with AS and MD. The study involved 36 patients with AS, 40 with MD, and 36 healthy controls (HCs). The participants were age-, sex-, and education-matched. Depression, anxiety, temperament, and character were evaluated using the Temperament and Character Inventory and physician-rated Hamilton Depression and Anxiety Scales. Disease severity was assessed using the Bath Ankylosing Spondylitis Disease Activity, Radiology, Metrology, and Functionality Indexes; erythrocyte sedimentation rate; and C-reactive protein. Depression and anxiety scores were highest in MD, lowest in HCs, and intermediate in AS (p < 0.001). Harm avoidance was significantly higher in MD and AS than HCs (p = 0.01). Reward dependency was lower (p = 0.011) and self-transcendence higher in AS only (p = 0.034). Bath Ankylosing Spondylitis Radiology Index was negatively correlated with depression (p = 0.020). Bath Ankylosing Spondylitis Radiology Index and Bath Ankylosing Spondylitis Metrology Index were negatively correlated with harm avoidance (p < 0.05). Disease duration was negatively correlated with self-transcendence (p = 0.027). Harm avoidance was correlated with depression score (p = 0.029). Patients under tumor necrosis factor α inhibitor treatment had lower cooperativeness score (p = 0.022). Patients with AS had a specific Temperament and Character Inventory pattern. Harm avoidance was common in both AS and MD. In AS, harm avoidance and depression decreased in patients with progression, probably because progressive fusion of the spinal segments causes decrease in pain. Harm avoidance (fear of pain) seemed to be an acting factor for occurrence of depressive symptoms in AS.

Rapid and quantitative in vitro analysis of mitochondrial fusion and its interplay with apoptosis

Mitochondrial fusion is essential to maintain genomic stability and physiological functions of mitochondria. Since mitochondrial fusion and fission work in concert to regulate mitochondrial morphology and functions, it has been challenging to quantitatively measure the direct roles of mitochondrial fusion in apoptosis and cancer progression. Here, we report the development of a high-throughput in vitro method to quantify mitochondrial fusion through single mitochondria analysis by a laboratory-built nano-flow cytometer (nFCM). Isolated mitochondria expressing green fluorescent protein (GFP-mito) or discosoma red fluorescent protein (DsRed-mito) were mixed together, induced to fuse, and analyzed by nFCM. A particle exhibiting both green and red fluorescence was identified as an event of heterotypic fusion, and the efficiency of heterotypic fusion was used as a surrogate of overall fusion efficiency. The as-developed method was applied to reveal the interplay between mitochondrial fusion and apoptosis without the interference of fission. We show that cytosolic components promoted mitochondrial fusion, and this upregulation was diminished during apoptosis. Combined with the translocation of Bid and Bax from cytosol to mitochondria, these findings suggest that cytosolic pro-apoptotic Bcl-2 family proteins could be the positive mediators of mitochondrial fusion. On the other hand, fusion also renders mitochondria more resistant to membrane potential collapse upon apoptosis induction. Our data suggest that disruption of mitochondrial fusion could be a potent strategy for cancer therapy. Furthermore, the as-developed method offers an effective approach to identify fusion inhibitors, including betulinic acid and antimycin A, giving reasons for their powerful utility in cancer treatment.

Deeply supervised convolutional neural network for shadow detection based on a novel aerial shadow imagery dataset

Shadow detection is an essential work for remote sensing image analysis, as the presence of shadows in high resolution images not only degrades the radiometric information but also disturbs the image interpretation. In this paper, a convolutional neural network (CNN) based shadow detection framework for aerial remote sensing images is presented. We construct a publicly available Aerial Imagery dataset for Shadow Detection (AISD), which is the first aerial shadow imagery dataset, as far as we know. Based on AISD, we propose a novel Deeply Supervised convolutional neural network for Shadow Detection (DSSDNet). To solve the insufficient feature extraction problem of shadows, the DSSDNet model is designed to include two steps: (1) an encoder-decoder residual (EDR) structure is adopted to extract multi-level and discriminative shadow features; (2) a deeply supervised progressive fusion (DSPF) process is then imposed on EDR to further boost the detection performance by directly guiding the training of the network and fuse adjacent feature maps progressively. The proposed DSSDNet is compared with several state-of-the-art methods in both qualitative and quantitative analysis. Results show that the proposed DSSDNet is more accurate, and more consistent to the shape of the objects casting shadows, with the average F-score being 91.79% on the testing images.

Enhanced mitochondrial fission suppresses signaling and metastasis in triple-negative breast cancer

BackgroundMitochondrial dynamics underlies malignant transformation, cancer progression, and response to treatment. Current research presents conflicting evidence for functions of mitochondrial fission and fusion in tumor progression. Here, we investigated how mitochondrial fission and fusion states regulate underlying processes of cancer progression and metastasis in triple-negative breast cancer (TNBC).MethodsWe enforced mitochondrial fission and fusion states through chemical or genetic approaches and measured migration and invasion of TNBC cells in 2D and 3D in vitro models. We also utilized kinase translocation reporters (KTRs) to identify single cell effects of mitochondrial state on signaling cascades, PI3K/Akt/mTOR and Ras/Raf/MEK/ERK, commonly activated in TNBC. Furthermore, we determined effects of fission and fusion states on metastasis, bone destruction, and signaling in mouse models of breast cancer.ResultsEnforcing mitochondrial fission through chemical or genetic approaches inhibited migration, invasion, and metastasis in TNBC. Breast cancer cells with predominantly fissioned mitochondria exhibited reduced activation of Akt and ERK both in vitro and in mouse models of breast cancer. Treatment with leflunomide, a potent activator of mitochondrial fusion proteins, overcame inhibitory effects of fission on migration, signaling, and metastasis. Mining existing datasets for breast cancer revealed that increased expression of genes associated with mitochondrial fission correlated with improved survival in human breast cancer.ConclusionsIn TNBC, mitochondrial fission inhibits cellular processes and signaling pathways associated with cancer progression and metastasis. These data suggest that therapies driving mitochondrial fission may benefit patients with breast cancer.

A Novel Human Placental Barrier Model Based on Trophoblast Stem Cells Derived from Human Induced Pluripotent Stem Cells.

The placenta acts as an interface between the fetus and the expecting mother. Various drugs and environmental pollutants can pass through the human placental barrier and may harm the developing fetus. Currently available in vitro placental barrier models are often inadequate, because they are lacking the functional trophoblast cells. Therefore, we developed and characterized a new human placental model using trophoblast stem cells (TSCs) derived from human induced pluripotent stem cells. Umbilical vein endothelial cells, fibroblast, and TSCs were cocultured using micromesh cell culture technique. These cells formed a tight three-layered structure. This coculture model induced progressive fusion of TSCs and formed a syncytialized epithelium that resembles the in vivo syncytiotrophoblast. Our model allowed the cultured trophoblasts to form microvilli and to reconstitute expression and physiological localization of membrane transport proteins, such as transporter for ATP-binding cassette subfamily B member 1, ATP-binding cassette subfamily C member 3, and glucose transporter-1. Drug permeability assays were performed using five compounds. The results from the permeability assays were comparable to the ones obtained with ex vivo placental models. In conclusion, we developed a novel coculture model mimicking human placenta that provides a useful tool for the studies on transfer of substances between the mother and fetus. Impact statement Compared with the currently available in vitro placental barrier models, a novel three-dimensional coculture placental barrier model presented in this study morphologically and functionally modeled the true placental barrier. The use of human trophoblast stem cells from human induced pluripotent stem cells substantially improved the current model. The use of micromesh sheet as a bioscaffold facilitated the formation of a good multilayer structure, which is closer to the physical appearance of the placenta observed in human.

Genome remodeling upon mesenchymal tumor cell fusion contributes to tumor progression and metastatic spread.

Cell fusion in tumor progression mostly refers to the merging of a cancer cell with a cell that has migration and immune escape capabilities such as macrophages. Here we show that spontaneous hybrids made from the fusion of transformed mesenchymal cells with partners from the same lineage undergo nonrecurrent large-scale genomic rearrangements, leading to the creation of highly aneuploid cells with novel phenotypic traits, including metastatic spreading capabilities. Moreover, in contrast to their parents, hybrids were the only cells able to recapitulate in vivo all features of human pleomorphic sarcomas, a rare and genetically complex mesenchymal tumor. Hybrid tumors not only displayed specific mesenchymal markers, but also combined a complex genetic profile with a highly metastatic behavior, like their human counterparts. Finally, we provide evidence that patient-derived pleomorphic sarcoma cells are inclined to spontaneous cell fusion. The resulting hybrids also gain in aggressiveness, exhibiting superior growth capacity in mouse models. Altogether, these results indicate that cell fusion has the potential to promote cancer progression by increasing growth and/or metastatic capacities, regardless of the nature of the companion cell. Moreover, such events likely occur upon sarcoma development, paving the way for better understanding of the biology, and aggressiveness of these tumors.

Multi-Person Pose Estimation Under Complex Environment Based on Progressive Rotation Correction and Multi-Scale Feature Fusion

The research of multi-person pose estimation has been largely improved recently. However, multi-person pose estimation in complex environments is still challenging. For example, the following two situations cannot be handled well by existing pose estimation methods: first, there are pedestrians that are not upright or even inverted in the image, and pedestrians of different scales appear in the same image. To solve these problems, the Progressive rotation correction module (PRCM) and Scale-invariance module (SIM) based on multi-scale feature fusion are proposed. First of all, the PRCM was proposed to address the situation where pedestrians appear rotated or even inverted in the image. This module is divided into three stages, with the aim of gradually correcting the inverted human to an upright one. Besides, SIM is designed to handle multi-scale problems. In this module, dilated convolutions with different receptive field are used to extract multi-scale information. Then, the extracted multi-scale features (different semantic information in different feature maps) will be fused to solve the multi-scale problem. The experimental results show that our algorithm can reach an AP value of 72.0% when tested on the COCO2017 dataset. Demonstrates that the proposed method is superior to state-of-the-art methods.