Auxiliary Branch Research Articles

The study on ultrasound image classification using a dual-branch model based on Resnet50 guided by U-net segmentation results

In recent years, the incidence of nodular thyroid diseases has been increasing annually. Ultrasonography has become a routine diagnostic tool for thyroid nodules due to its high real-time capabilities and low invasiveness. However, thyroid images obtained from current ultrasound tests often have low resolution and are plagued by significant noise interference. Regional differences in medical conditions and varying levels of physician experience can impact the accuracy and efficiency of diagnostic results. With the advancement of deep learning technology, deep learning models are used to identify whether a nodule in a thyroid ultrasound image is benign or malignant. This helps to close the gap between doctors’ experience and equipment differences, improving the accuracy of the initial diagnosis of thyroid nodules. To cope with the problem that thyroid ultrasound images contain complex background and noise as well as poorly defined local features. in this paper, we first construct an improved ResNet50 classification model that uses a two-branch input and incorporates a global attention lightening module. This model is used to improve the accuracy of benign and malignant nodule classification in thyroid ultrasound images and to reduce the computational effort due to the two-branch structure.We constructed a U-net segmentation model incorporating our proposed ACR module, which uses hollow convolution with different dilation rates to capture multi-scale contextual information for feature extraction of nodules in thyroid ultrasound images and uses the results of the segmentation task as an auxiliary branch for the classification task to guide the classification model to focus on the lesion region more efficiently in the case of weak local features. The classification model is guided to focus on the lesion region more efficiently, and the classification and segmentation sub-networks are respectively improved specifically for this study, which is used to improve the accuracy of classifying the benign and malignant nature of the nodules in thyroid ultrasound images. The experimental results show that the four evaluation metrics of accuracy, precision, recall, and f1 of the improved model are 96.01%, 93.3%, 98.8%, and 96.0%, respectively. The improvements were 5.7%, 1.6%, 13.1%, and 7.4%, respectively, compared with the baseline classification model.

MAR-YOLOv9: A multi-dataset object detection method for agricultural fields based on YOLOv9

With the development of deep learning technology, object detection has been widely applied in various fields. However, in cross-dataset object detection, conventional deep learning models often face performance degradation issues. This is particularly true in the agricultural field, where there is a multitude of crop types and a complex and variable environment. Existing technologies still face performance bottlenecks when dealing with diverse scenarios. To address these issues, this study proposes a lightweight, cross-dataset enhanced object detection method for the agricultural domain based on YOLOv9, named Multi-Adapt Recognition-YOLOv9 (MAR-YOLOv9). The traditional 32x downsampling Backbone network has been optimized, and a 16x downsampling Backbone network has been innovatively designed. A more streamlined and lightweight Main Neck structure has been introduced, along with innovative methods for feature extraction, up-sampling, and Concat connection. The hybrid connection strategy allows the model to flexibly utilize features from different levels. This solves the issues of increased training time and redundant weights caused by the detection neck and auxiliary branch structures in traditional YOLOv9, enabling MAR-YOLOv9 to maintain high performance while reducing the model’s computational complexity and improving detection speed, making it more suitable for real-time detection tasks. In comparative experiments on four plant datasets, MAR-YOLOv9 improved the mAP@0.5 accuracy by 39.18% compared to seven mainstream object detection algorithms, and by 1.28% compared to the YOLOv9 model. At the same time, the model size was reduced by 9.3%, and the number of model layers was decreased, reducing computational costs and storage requirements. Additionally, MAR-YOLOv9 demonstrated significant advantages in detecting complex agricultural images, providing an efficient, lightweight, and adaptable solution for object detection tasks in the agricultural field. The curated data and code can be accessed at the following link: https://github.com/YangxuWangamI/MAR-YOLOv9.

MW‐SAM:Mangrove wetland remote sensing image segmentation network based on segment anything model

AbstractMangrove wetlands are important ecosystems in tropical and subtropical coastal areas, providing wind and wave attenuation and embankment protection functions. However, mangrove wetlands worldwide are facing severe loss and degradation. Accurate identification of mangrove wetland extent is crucial for their protection, but traditional methods struggle to meet the requirements of large‐scale, high‐precision identification. Furthermore, field data collection and annotation in wetlands in complex intertidal zones pose challenges, and the lack of high‐quality training samples limits the application of deep learning methods in this field. This paper proposes a novel semantic segmentation framework for mangrove wetlands called mangrove wetland remote sensing image segmentation network based on segment anything model (MW‐SAM) to address these issues. MW‐SAM is based on the pre‐trained SAM and achieves cross‐domain adaptation through parameter‐efficient fine‐tuning techniques. It introduces wetland‐specific prompts, auxiliary branches, semi‐supervised training, and iterative optimization strategies to improve accuracy and tackle the problem of sample scarcity. Moreover, on the mangrove wetland dataset constructed in this paper, MW‐SAM significantly outperforms SAM and other traditional methods. MW‐SAM provides a new technology for monitoring and protecting mangrove wetlands, and the constructed dataset will be made publicly available, which is expected to promote the development of mangrove wetland conservation efforts.

Driver behavior recognition based on dual-branch and deformable convolutional network method

Aiming at the task of driver behavior recognition in the car cockpit, this paper proposes a recognition method based on a dual-branch neural network. The main branch of the network model uses ResNet50 as the backbone network for feature extraction, and uses deformable convolution to adapt the model to the shape and position changes of the driver in the image. The auxiliary branch assists in updating the parameters of the backbone network during the gradient backpropagation process, so that the backbone network can better extract features that are conducive to driver behavior recognition, thereby improving the recognition performance of the model. The ablation experiment and comparative experiment results of the network model on the State Farm public dataset show that the recognition accuracy of the proposed network model can reach 96.23%, and the recognition effect is better for easily confused behavior categories. The research results are of great significance for understanding driver behavior in the car cockpit and ensuring driving safety.

Herschel-Quincke filters adapted to bending waves in beams and plates

A Herschel-Quincke (HQ) filter is a device for controlling the transmission of guided waves: by adding one or more auxiliary branches to a waveguide, it is possible to construct delay lines, which induce destructive interference effects and consequently zero-transmission phenomena at targeted frequencies. We explore the use of this principle in the context of elastic waves in beams.The dispersive nature of bending waves and their coupling with longitudinal waves introduces complexity into the tuning of the zero transmission frequencies. A model, developed within the framework of Timoshenko's beam theory, provides the scattering matrix of an HQ filter composed of several branches. The model is confirmed by FE and thus demonstrates the interest of the approach for the control of vibrations. The architecture of the HQ device and the modelling approach are extended to a 3 branches configuration. The properties of the filter are analysed, leading to a discussion of the advantages and disadvantages of the HQ strategy.

A dual branch graph neural network based spatial interpolation method for traffic data inference in unobserved locations

Complete traffic data collection is crucial for intelligent transportation system, but due to various factors such as cost, it is not possible to deploy sensors at every location. Using spatial interpolation, the traffic data for unobserved locations can be inferred from the data of observed locations, providing fine-grained measurements for improved traffic monitoring and control. However, existing methods are limited in modeling the dynamic spatio-temporal dependencies between traffic locations, resulting in unsatisfactory performance of spatial interpolation for unobserved locations in traffic scene. To address this issue, we propose a novel dual branch graph neural network (DBGNN) for spatial interpolation by exploiting dynamic spatio-temporal correlation among traffic nodes. The proposed DBGNN is composed of two branches: the main branch and the auxiliary branch. They are designed to capture the wide-range dynamic spatial correlation and the local detailed spatial diffusion between nodes, respectively. Finally, they are fused via a self-attention mechanism. Extensive experiments on six public datasets demonstrate the advantages of our DBGNN over the state-of-the-art baselines. The codes will be available at https://github.com/SYLan2019/DBGNN.

LDTR: Transformer-based lane detection with anchor-chain representation

Despite recent advances in lane detection methods, scenarios with limited- or no-visual-clue of lanes due to factors such as lighting conditions and occlusion remain challenging and crucial for automated driving. Moreover, current lane representations require complex post-processing and struggle with specific instances. Inspired by the DETR architecture, we propose LDTR, a transformer-based model to address these issues. Lanes are modeled with a novel anchor-chain, regarding a lane as a whole from the beginning, which enables LDTR to handle special lanes inherently. To enhance lane instance perception, LDTR incorporates a novel multi-referenced deformable attention module to distribute attention around the object. Additionally, LDTR incorporates two line IoU algorithms to improve convergence efficiency and employs a Gaussian heatmap auxiliary branch to enhance model representation capability during training. To evaluate lane detection models, we rely on Fréchet distance, parameterized Fl-score, and additional synthetic metrics. Experimental results demonstrate that LDTR achieves state-of-the-art performance on well-known datasets.

D2GL: Dual-level dual-scale graph learning for sketch-based 3D shape retrieval

Sketch-based 3D shape retrieval (SBSR) is an active research area in the computer vision community, but it is still very challenging. One main reason is that existing deep learning-based methods usually treat sketches as 2D images, neglecting the sparsity and diversity. In this paper, we propose a novel Dual-level Dual-scale Graph Learning (D2GL) method to effectively enhance structural information and produce robust representations for sparse and diverse hand-drawn sketches. Specifically, in addition to the traditional branches for SBSR, we introduce a Dual-level Dual-scale Graph Self-attention (DLDS-GSA) as an auxiliary branch. DLDS-GSA further consists of two levels of encoders, i.e., a local structural encoder and a dual-scale global structural encoder, to capture both local discriminative and multi-scale global structures while minimizing the impact of various sketch drawing details. Comprehensive experiments on SHREC’13 and SHREC’14 datasets demonstrate the superiority of D2GL for SBSR, with extended experiments on PART-SHREC’14 confirming its generalization for unseen classes in SBSR.

TMP: Temporal Motion Perception with spatial auxiliary enhancement for moving Infrared dim-small target detection

Infrared dim and small target detection has significant research value and application prospects in both military affairs and civil safety scenes. This topic presents substantial challenges due to the small size of the targets and their low contrast against the background in images. In the past ten years, most detection methodologies have primarily focused on single-frame images, mainly relying on spatial-domain image features to identify dim and small targets, with limited consideration of temporal-domain motion features. However, unlike general objects, the image features of dim-small targets are often easy to be interfered by background, and even seriously lost in capturing stages. These deficiencies can be the primary factors to impact the detection performance of single-frame methods. To promote infrared dim-small target detection, this paper proposes a novel multi-frame pipeline called Temporal Motion Perception with spatial auxiliary enhancement (TMP), which consists of two parallel feature extraction branches. One auxiliary branch is designed to capture traditional image features from spatial domain, and the other is proposed to extract the motion features of targets from temporal domain. In addition, a new Complementary Symmetry Weighting module is specially designed to fulfill the cross-domain fusion of spatio-temporal features, further enhancing feature representation. Experimental results on three public datasets demonstrate the efficiency of our proposed scheme. Compared with existing methods, it could often refresh the state-of-the-art (SOTA) performance on most metrics. Our codes are available at https://github.com/UESTC-nnLab/TMP.

A Faster and Lightweight Lane Detection Method in Complex Scenarios

Lane detection is a crucial visual perception task in the field of autonomous driving, serving as one of the core modules in advanced driver assistance systems (ADASs).To address the insufficient real-time performance of current segmentation-based models and the conflict between the demand for high inference speed and the excessive parameters in resource-constrained edge devices (such as onboard hardware, mobile terminals, etc.) in complex real-world scenarios, this paper proposes an efficient and lightweight auxiliary branch network (CBGA-Auxiliary) to tackle these issues. Firstly, to enhance the model’s capability to extract feature information in complex scenarios, a row anchor-based feature extraction method based on global features was adopted. Secondly, employing ResNet as the backbone network and CBGA (Conv-Bn-GELU-SE Attention) as the fundamental module, we formed the auxiliary segmentation network, significantly enhancing the segmentation training speed of the model. Additionally, we replaced the standard convolutions in the branch network with lightweight GhostConv convolutions. This reduced the parameters and computational complexity while maintaining accuracy. Finally, an additional enhanced structural loss function was introduced to compensate for the structural defect loss issue inherent in the row anchor-based method, further improving the detection accuracy. The model underwent extensive experimentation on the Tusimple dataset and the CULane dataset, which encompass various road scenarios. The experimental results indicate that the model achieved the highest F1 scores of 96.1% and 71.0% on the Tusimple and CULane datasets, respectively. At a resolution of 288 × 800, the ResNet18 and ResNet34 models achieved maximum inference speeds of 410FPS and 280FPS, respectively. Compared to existing SOTA models, it demonstrates a significant advantage in terms of inference speed. The model achieved a good balance between accuracy and inference speed, making it suitable for deployment on edge devices and validates the effectiveness of the model.

Landslide extraction from aerial imagery considering context association characteristics

Accurate extraction of landslide information is crucial for timely disaster emergency response, yet this process faces significant challenges due to the interference of bare land and vegetation cover as well as the variability in landslide scales. Furthermore, the scarcity of comprehensive landslide datasets has hindered the in-depth exploration of deep learning techniques in this domain. Addressing these issues, this study introduces a novel framework for landslide extraction that leverages the characteristics of context association. In the encoding phase, a two-branch multiscale context feature extraction module (TMCFM) is established, which captures the contextual relationships across different scales through an attention mechanism while concurrently extracting context information within the same scale. To refine the interplay between features at different levels, self-attention is utilized, allowing for the effective fusion of context information at both ends of the same scale through feature fusion exploiting static and dynamic context feature modules (FSDC). Building upon this, this study develops a deeply supervised classifier (DSC) that enhances the network’s discriminative capabilities in the prediction phase via six auxiliary branches. Additionally, this study contributes a new aerial image dataset for landslide extraction created through meticulous visual interpretation. The proposed method is thoroughly compared with 17 contemporary deep learning methods, demonstrating an increase in the intersection-over-union (IoU) metric of 0.92%-16.94% over these methods. The robustness and superiority of the proposed approach are further validated through various discussions and analyses.

Identification of Radio Frequency Interference Using Multi-scale TransUNet

Radio observation is a method for conducting astronomical observations using radio waves. A common challenge in radio observations is Radio Frequency Interference (RFI), which refers to the unintentional or intentional interference of radio signals from other wireless sources within the radio frequency band. Such interference contaminates the astronomical signals received by radio telescopes, significantly affecting time–frequency domain astronomical observations and research. Consequently, identifying RFI is crucial. In this paper, we employ a deep learning approach to detect RFI present in observation data and propose an improved network structure based on TransUNet. This network leverages the principles of a multi-scale convolutional attention mechanism. It introduces an auxiliary branch to extract high-dimensional image information and an enhanced coordinate attention mechanism for feature map extraction, enabling more comprehensive and accurate identification of RFI in time–frequency images. We introduce a novel architecture named the Multi-Scale TransUNet Network, abbreviated as MS-TransUNet. We utilized observation data from the 40 m radio telescope at the Yunnan Observatory as a data set for training, validating, and testing the network. Compared with previous deep learning networks (U-Net, RFI-Net, R-Net, DSC, EMSCA-UNet), the recall rate and f2 score have been significantly improved. Specifically, the recall rate is improved by at least 2.99%, and the f2 score is improved by at least 2.46%. Experiments demonstrate that this network is exceptional in identifying RFI more comprehensively while ensuring high precision.

A mm‐wave LNA employing current re‐use and non‐linearity cancellation in 28 nm CMOS for automotive RADAR and 6G receivers

AbstractThis letter reports an 85 GHz low noise amplifier (LNA) employing derivative superposition based non‐linearity cancellation and a current re‐use topology. The LNA employs a two‐stage stacked architecture, each featuring neutralized differential pairs utilizing the same DC current. In derivative superposition, an auxiliary branch consisting of neutralized differential pairs cells is added in the LNA in parallel to stage 1 to ensure non‐linearity cancellation. Layout‐based capacitive neutralization is implemented to improve GMAX, resulting in simplified routing, reduced parasitics, and a more compact layout. The proposed LNA is fabricated in TSMC 28 nm CMOS process and achieves a peak gain of 8.4 dB at 84.2 GHz with a measured 3 dB bandwidth () from 79.4 to 92 GHz. The minimum measured noise figure is 12.8 dB. The LNA draws 34 mA of DC current from a 1.2 V supply. The highly linear LNA with IIP3 +6.5 dBm is tailored for automotive RADAR and 6G receivers.

Learning Complementary Correlations for Depth Super-Resolution With Incomplete Data in Real World.

Depth information is a significant ingredient to visually perceive the physical world. However, mainstream depth sensors, e.g., time-of-flight (ToF) cameras, often measure incomplete and low-resolution depth data, resulting in low-quality visual perception. In this article, we try to address a potentially valuable task, i.e., depth super-resolution (DSR) with incomplete data, which recovers dense and high-resolution depth map from incomplete and low-resolution one. To tackle this task, we introduce a novel incomplete DSR (IDSR) framework, including a primary branch for DSR to recover high-frequency details, and an auxiliary branch for depth completion (DC) to fill missing pixels. More importantly, we propose two modules, joint correlation learning (JCL) and iterative-cross (IC), to enhance the learning of complementary information flows between the two branches. The former module aims to learn the correlative relationships of the two branches, whilst the latter module adequately fuses higher level representations for more precise predictions. Extensive experiments show that our framework is effective and achieves the state-of-the-art performance on the real-world RGB-D-D and the synthetic NYUv2 datasets.

Novel Series-Parallel Phase-Shifted Full-Bridge Converters with Auxiliary LC Networks to Achieve Wide Lagging-Leg ZVS Range

Under light load conditions, the phase-shifted full-bridge (PSFB) converter often has difficulty in realizing the zero-voltage switching (ZVS) of the lagging-leg by relying on the energy of its resonant inductor; however, for the series-parallel PSFB converter applied in high-power applications, the lagging-leg still has the problem of difficult realization of ZVS. Based on this, the paper analyzes the reasons why the series-parallel PSFB converter has difficulty in achieving ZVS for the lagging-leg under light and heavy loads. Under interleaved control, the ZVS of the lagging-leg over the full load range is realized by adding an auxiliary LC branch at the midpoint of the lagging-leg of both submodules. Based on the double-bridge input-parallel-output-series (IPOS) PSFB converter, analyzing the working principle of the circuit after adding the auxiliary LC branch and extending it to the series-parallel PSFB converter. The design requirements of the LC auxiliary branch of the dual-bridge series-parallel PSFB converter are given and the effects of the LC auxiliary branch on the module operating state and device stress are analyzed. On this basis, an extension is carried out to give the working principle and design method of the auxiliary LC branch of the N-bridge series-parallel PSFB converter. Finally, a 100 kW Matlab/Simulink simulation model verifies the superior performance of the proposed LC auxiliary branch to realize the lagging-leg ZVS of the series-parallel PSFB converter under light and heavy loads and achieves a 1.09% peak efficiency improvement at rated load.

Face anti-spoofing with cross-stage relation enhancement and spoof material perception

Face Anti-Spoofing (FAS) seeks to protect face recognition systems from spoofing attacks, which is applied extensively in scenarios such as access control, electronic payment, and security surveillance systems. Face anti-spoofing requires the integration of local details and global semantic information. Existing CNN-based methods rely on small stride or image patch-based feature extraction structures, which struggle to capture spatial and cross-layer feature correlations effectively. Meanwhile, Transformer-based methods have limitations in extracting discriminative detailed features. To address the aforementioned issues, we introduce a multi-stage CNN-Transformer-based framework, which extracts local features through the convolutional layer and long-distance feature relationships via self-attention. Based on this, we proposed a cross-attention multi-stage feature fusion, employing semantically high-stage features to query task-relevant features in low-stage features for further cross-stage feature fusion. To enhance the discrimination of local features for subtle differences, we design pixel-wise material classification supervision and add a auxiliary branch in the intermediate layers of the model. Moreover, to address the limitations of a single acquisition environment and scarcity of acquisition devices in the existing Near-Infrared dataset, we create a large-scale Near-Infrared Face Anti-Spoofing dataset with 380k pictures of 1040 identities. The proposed method could achieve the state-of-the-art in OULU-NPU and our proposed Near-Infrared dataset at just 1.3GFlops and 3.2M parameter numbers, which demonstrate the effective of the proposed method.

A Fast and Accurate Lane Detection Method Based on Row Anchor and Transformer Structure.

Lane detection plays a pivotal role in the successful implementation of Advanced Driver Assistance Systems (ADASs), which are essential for detecting the road's lane markings and determining the vehicle's position, thereby influencing subsequent decision making. However, current deep learning-based lane detection methods encounter challenges. Firstly, the on-board hardware limitations necessitate an exceptionally fast prediction speed for the lane detection method. Secondly, improvements are required for effective lane detection in complex scenarios. This paper addresses these issues by enhancing the row-anchor-based lane detection method. The Transformer encoder-decoder structure is leveraged as the row classification enhances the model's capability to extract global features and detect lane lines in intricate environments. The Feature-aligned Pyramid Network (FaPN) structure serves as an auxiliary branch, complemented by a novel structural loss with expectation loss, further refining the method's accuracy. The experimental results demonstrate our method's commendable accuracy and real-time performance, achieving a rapid prediction speed of 129 FPS (the single prediction time of the model on RTX3080 is 15.72 ms) and a 96.16% accuracy on the Tusimple dataset-a 3.32% improvement compared to the baseline method.

Test-Time Domain Adaptation by Learning Domain-Aware Batch Normalization

Test-time domain adaptation aims to adapt the model trained on source domains to unseen target domains using a few unlabeled images. Emerging research has shown that the label and domain information is separately embedded in the weight matrix and batch normalization (BN) layer. Previous works normally update the whole network naively without explicitly decoupling the knowledge between label and domain. As a result, it leads to knowledge interference and defective distribution adaptation. In this work, we propose to reduce such learning interference and elevate the domain knowledge learning by only manipulating the BN layer. However, the normalization step in BN is intrinsically unstable when the statistics are re-estimated from a few samples. We find that ambiguities can be greatly reduced when only updating the two affine parameters in BN while keeping the source domain statistics. To further enhance the domain knowledge extraction from unlabeled data, we construct an auxiliary branch with label-independent self-supervised learning (SSL) to provide supervision. Moreover, we propose a bi-level optimization based on meta-learning to enforce the alignment of two learning objectives of auxiliary and main branches. The goal is to use the auxiliary branch to adapt the domain and benefit main task for subsequent inference. Our method keeps the same computational cost at inference as the auxiliary branch can be thoroughly discarded after adaptation. Extensive experiments show that our method outperforms the prior works on five WILDS real-world domain shift datasets. Our method can also be integrated with methods with label-dependent optimization to further push the performance boundary. Our code is available at https://github.com/ynanwu/MABN.

Two-stage robust optimization of unified power quality conditioner (UPQC) siting and sizing in active distribution networks considering uncertainty of loads and renewable generators

This paper proposes a robust optimization siting and sizing strategy for the unified power quality conditioner (UPQC) in the active distribution network (ADN) integrated with renewable energy generators. The optimization objectives including investment cost, line loss, and voltage deviation are integrated into an overall cost for both the planning and operating stages. A novel UPQC equivalent power injection model is proposed based on the auxiliary branch method. A set of linear and second-order conic (SOC) constraints are employed to accurately describe the steady-state operation of UPQC. An improved branch current flow model is developed to incorporate radial distribution networks with UPQCs. The UPQC siting and sizing planning is formulated as a mixed integer second-order conic programming (MISOCP) problem. Accounting for the uncertainty in loads and distributed generators, a two-stage robust optimization model is established to address worst-case operational scenarios. A solving strategy is developed based on the strong duality theory and the column-and-constraint generation (C&CG) algorithm, and then commercial solvers can solve the problem accurately and efficiently. Case studies on an IEEE 33-node distribution network demonstrate this strategy's robustness in dealing with loads and renewable energy uncertainties, making it applicable across different UPQC topologies and operation modes. The proposed method has superior performance over the representative meta-heuristic approaches in terms of global optimality as well as an over 7.5 times faster computation speed. This work has significance in enhancing the quality and efficiency of modern power systems highly integrated with renewable energy.

Knowledge Distillation Using Hierarchical Self-Supervision Augmented Distribution.

Knowledge distillation (KD) is an effective framework that aims to transfer meaningful information from a large teacher to a smaller student. Generally, KD often involves how to define and transfer knowledge. Previous KD methods often focus on mining various forms of knowledge, for example, feature maps and refined information. However, the knowledge is derived from the primary supervised task, and thus, is highly task-specific. Motivated by the recent success of self-supervised representation learning, we propose an auxiliary self-supervision augmented task to guide networks to learn more meaningful features. Therefore, we can derive soft self-supervision augmented distributions as richer dark knowledge from this task for KD. Unlike previous knowledge, this distribution encodes joint knowledge from supervised and self-supervised feature learning. Beyond knowledge exploration, we propose to append several auxiliary branches at various hidden layers, to fully take advantage of hierarchical feature maps. Each auxiliary branch is guided to learn self-supervision augmented tasks and distill this distribution from teacher to student. Overall, we call our KD method a hierarchical self-supervision augmented KD (HSSAKD). Experiments on standard image classification show that both offline and online HSSAKD achieves state-of-the-art performance in the field of KD. Further transfer experiments on object detection further verify that HSSAKD can guide the network to learn better features. The code is available at https://github.com/winycg/HSAKD.