Convolution Kernel Research Articles

MSMP-Net: A Multi-Scale Neural Network for End-to-End Monkeypox Virus Skin Lesion Classification

Monkeypox is a zoonotic disease caused by monkeypox virus infection. It is easily transmitted among people and poses a major threat to human health, making it of great significance in public health. Therefore, this paper proposes MSMP-Net, a multi-scale neural network for end-to-end monkeypox virus skin lesion classification ConvNeXt is used as the backbone network, and designs such as inverse bottleneck layers and large convolution kernels are used to enhance the network’s feature extraction capabilities. In order to effectively utilize the multi-level feature maps generated by the backbone network, a multi-scale feature fusion structure was designed. By fusing the deepest feature maps of multi-scale features, the model’s ability to represent monkeypox image features is enhanced. Experimental results show that the accuracy, precision, recall, and F1-score of this method on the MSLD v2.0 dataset are 87.03 ± 3.43%, 87.59 ± 3.37%, 87.03 ± 3.43%, and 86.58 ± 3.66%, respectively.

A Ship’s Maritime Critical Target Identification Method Based on Lightweight and Triple Attention Mechanisms

The ability to classify and recognize maritime targets based on visual images plays an important role in advancing ship intelligence and digitalization. The current target recognition algorithms for common maritime targets, such as buoys, reefs, other ships, and bridges of different colors, face challenges such as incomplete classification, low recognition accuracy, and a large number of model parameters. To address these issues, this paper proposes a novel maritime target recognition method called DTI-YOL (DualConv Triple Attention InnerEIOU-You Only Look Once). This method is based on a triple attention mechanism designed to enhance the model’s ability to classify and recognize buoys of different colors in the channel while also making the feature extraction network more lightweight. First, the lightweight double convolution kernel feature extraction layer is constructed using group convolution technology to replace the Conv structure of YOLOv9 (You Only Look Once Version 9), effectively reducing the number of parameters in the original model. Second, an improved three-branch structure is designed to capture cross-dimensional interactions of input image features. This structure forms a triple attention mechanism that accounts for the mutual dependencies between input channels and spatial positions, allowing for the calculation of attention weights for targets such as bridges, buoys, and other ships. Finally, InnerEIoU is used to replace CIoU to improve the loss function, thereby optimizing loss regression for targets with large scale differences. To verify the effectiveness of these algorithmic improvements, the DTI-YOLO algorithm was tested on a self-made dataset of 2300 ship navigation images. The experimental results show that the average accuracy of this method in identifying seven types of targets—including buoys, bridges, islands and reefs, container ships, bulk carriers, passenger ships, and other ships—reached 92.1%, with a 12% reduction in the number of parameters. This enhancement improves the model’s ability to recognize and distinguish different targets and buoy colors.

Importance of OCT-derived biomarkers for the recurrence of central serous chorioretinopathy using statistics and predictive modelling

Central serous chorioretinopathy (CSCR) is a retinal disease characterised by the accumulation of subretinal fluid, which often resolves spontaneously in acute cases. However, approximately one-third of patients experience recurrences that may cause severe and irreversible vision. This study aimed to identify parameters derived from optical coherence tomography (OCT) that are associated with CSCR recurrence. Our dataset included 5211 OCT scans from 344 eyes of 255 patients diagnosed with CSCR. 178 eyes were identified as recurrent, 109 as non-recurrent, and 57 were excluded. We extracted parameters using artificial intelligence algorithms based on U-Nets, convolutional kernels, and morphological operators. We applied inferential statistics to evaluate differences between the recurrent and non-recurrent groups, and we used a logistic regression predictive model, reporting the coefficients as a measure of biomarker importance. We identified nine predictive biomarkers for CSCR recurrence: age, intraretinal fluid, subretinal fluid, pigment epithelial detachments, choroidal vascularity index, integrity of photoreceptors and retinal pigment epithelium layer, choriocapillaris and choroidal stroma thickness, and thinning of the outer nuclear layer, and of the inner nuclear layer combined with the outer plexiform layer. These results could enable future developments in the automatic detection of CSCR recurrence, paving the way for translational medical applications.

High‐Resolution Optical Convolutional Neural Networks Using Phase‐Change Material‐Based Microring Hybrid Waveguides

In the More‐than‐Moore era, the explosive growth of data and information has driven the exploration of alternative non‐von Neumann computational paradigms. Photonic neuromorphic computing has emerged as a promising approach, offering high speed, wide bandwidth, and massive parallelism. Herein, a high‐resolution optical convolutional neural network (OCNN) is introduced using phase‐change material Ge2Sb2Te5 (GST)‐based microring hybrid waveguides. This on‐chip optical computing platform integrates GST into photonic devices, enabling versatile programming and in‐memory computing capabilities. Central to this platform is a photonic convolutional computational kernel, constructed from photonic switching cells embedded with GST on a microring resonator. This programmable photonic switch leverages the refractive index modulation during the GST phase transition to achieve up to 64 discrete levels of transmission contrast, suitable for representing matrix elements in neural network algorithms with 6‐bit resolution. Using these matrix elements, an OCNN capable of performing parallelized image edge detection and digital recognition tasks with high accuracy is demonstrated. The architecture is scalable for large‐scale photonic neural networks, offering ultrahigh computational throughput, a compact design, complementary metal‐oxide‐semiconductor‐compatible fabrication, and broad bandwidth.

Multi-Source Information-Based Bearing Fault Diagnosis Using Multi-Branch Selective Fusion Deep Residual Network.

Rolling bearing is the core component of industrial machines, but it is difficult for common single signal source-based fault diagnosis methods to ensure reliable results since sensor signals are vulnerable to the pollution of background noises and the attenuation of transmitted information. Recently, multi-source information-based fault diagnosis methods have become popular, but the information redundancy between multiple signals is a tough problem that will negatively impact the representational capacity of deep learning algorithms and the precision of fault diagnosis methods. Besides that, the characteristics of various signals are actually different, but this problem was usually omitted by researchers, and it has potential to further improve the diagnosing performance by adaptively adjusting the feature extraction process for every input signal source. Aimed at solving the above problems, a novel model for bearing fault diagnosis called multi-branch selective fusion deep residual network is proposed in this paper. The model adopts a multi-branch structure design to enable every input signal source to have a unique feature processing channel, avoiding the information of multiple signal sources blindly coupled by convolution kernels. And in each branch, different convolution kernel sizes are assigned according to the characteristics of every input signal, fully digging the precious fault components on respective information sources. Lastly, the dropout technique is used to randomly throw out some activated neurons, alleviating the redundancy and enhancing the quality of the multiscale features extracted from different signals. The proposed method was experimentally compared with other intelligent methods on two authoritative public bearing datasets, and the experimental results prove the feasibility and superiority of the proposed model.

Speckle-free holography with a diffraction-aware global perceptual model

Computer-generated holography (CGH) based on neural networks has been actively investigated in recent years, and convolutional neural networks (CNNs) are frequently adopted. A convolutional kernel captures local dependencies between neighboring pixels. However, in CGH, each pixel on the hologram influences all the image pixels on the observation plane, thus requiring a network capable of learning long-distance dependencies. To tackle this problem, we propose a CGH model called Holomer. Its single-layer perceptual field is 43 times larger than that of a widely used 3×3 convolutional kernel, thanks to the embedding-based feature dimensionality reduction and multi-head sliding-window self-attention mechanisms. In addition, we propose a metric to measure the networks’ learning ability of the inverse diffraction process. In the simulation, our method demonstrated noteworthy performance on the DIV2K dataset at a resolution of 1920×1024, achieving a PSNR and an SSIM of 35.59 dB and 0.93, respectively. The optical experiments reveal that our results have excellent image details and no observable background speckle noise. This work paves the path of high-quality hologram generation.

A novel embedded kernel CNN-PCFF algorithm for breast cancer pathological image classification

Early screening of breast cancer through image recognition technology can significantly increase the survival rate of patients. Therefore, breast cancer pathological image is of great significance for medical diagnosis and clinical research. In recent years, numerous deep learning models have been applied to breast cancer image classification, with deep CNN being a typical representative. Due to the use of multi-depth small convolutional kernels in mainstream CNN architectures such as VGG and Inception, the obtained image features often have high dimensionality. Although high dimensionality can bring more fine-grained features, it also increases the computational complexity of subsequent classifiers and may even lead to the curse of dimensionality and overfitting. To address these issues, a novel embedded kernel CNN principal component feature fusion (CNN-PCFF) algorithm is proposed. The constructed kernel function is embedded in the principal component analysis to form the multi-kernel principal component. Multi-kernel principal component analysis is used to fuse the high dimensional features obtained from the convolution base into some representative comprehensive variables, which are called kernel principal components, so as to achieve the purpose of dimensionality reduction. Any type of classifier can be added based on multi-kernel principal components. Through experimental analysis on two public breast cancer image datasets, the results show that the proposed algorithm can improve the performance of the current mainstream CNN architecture and subsequent classifiers. Therefore, the proposed algorithm in this paper is an effective tool for the classification of breast cancer pathological images.

BIF-Net: Boundary information fusion network for abdominal aortic aneurysm segmentation

The accurate abdominal aortic aneurysm (AAA) segmentation is significant for assisting clinicians in diagnosis and treatment planning. However, existing segmentation methods exhibit a low utilization rate for the semantic information of vessel boundaries, which is disadvantageous for segmenting AAA with significant scale variability of vessel diameter (diameter ranges from 4 mm to 85 mm). To tackle this problem, we introduce a boundary information fusion network (BIF-Net) specially designed for AAA segmentation. BIF-Net initially constructs convolutional kernel functions based on Gabor and Sobel operators, enriching the global semantic features and localization information through the Gabor and Sobel dilated convolution (GSDC) module. Additionally, BIF-Net supplements lost boundary feature information during the sampling process through the guided filtering feature supplementation (GFFS) module and the channel-spatial attention module (CSAM), enhancing the ability to capture targets with shape diversity and boundary features. Finally, we introduce a boundary feature loss function to alleviate the impact of the imbalance between positive and negative samples. The results demonstrate that BIF-Net outperforms current state-of-the-art methods across multiple evaluation metrics, achieving the highest Dice similarity coefficient (DSC) accuracies of 93.29 % and 91.01 % on the preoperative and postoperative datasets, respectively. Compared to the state-of-the-art methods, BIF-Net improves DSC by 6.86 % and 3.85 %. Due to the powerful boundary feature extraction ability, the proposed BIF-Net is a competitive AAA segmentation method exhibiting significant potential for application in diagnosis and treatment processes of AAA.

Breaking the rain barrier: A novel approach in image processing with AMGR‐Net

AbstractImage quality is significantly impacted by rain, posing challenges in fields like surveillance, autonomous driving, and outdoor robotics. The field of image deraining, particularly for single image, has attracted considerable attention to improve image clarity in inclement weather. To overcome the inherent complexity of the single‐image rain removal task, we proposed a novel architecture of the attention mechanism and gated recurrent network (AMGR‐Net) that combines spatial and channel attention mechanisms with gated recurrent units. AMGR‐Net contains multiple modules, each of which uses convolution kernels and attention mechanisms to enhance feature extraction. AMGR‐Net demonstrates superior performance over state‐of‐the‐art methods in both synthetic and real‐world image datasets, as evidenced by higher peak signal to noise ratio and structural similarity index measurement scores. The integration of spatial attention significantly enhances feature expression, enabling more effective rain streak removal and detail preservation. Furthermore, this method also shows promising results in the application of stripe noise removal from meteorological satellite cloud images.

A semantic segmentation method integrated convolutional nonlinear spiking neural model with Transformer

Semantic segmentation is a critical task in computer vision, with significant applications in areas like autonomous driving and medical imaging. Transformer-based methods have gained considerable attention recently because of their strength in capturing global information. However, these methods often sacrifice detailed information due to the lack of mechanisms for local interactions. Similarly, convolutional neural network (CNN) methods struggle to capture global context due to the inherent limitations of convolutional kernels. To overcome these challenges, this paper introduces a novel Transformer-based semantic segmentation method called NSNPFormer, which leverages the nonlinear spiking neural P (NSNP) system—a computational model inspired by the spiking mechanisms of biological neurons. The NSNPFormer employs an encoding–decoding structure with two convolutional NSNP components and a residual connection channel. The convolutional NSNP components facilitate nonlinear local feature extraction and block-level feature fusion. Meanwhile, the residual connection channel helps prevent the loss of feature information during the decoding process. Evaluations on the ADE20K and Pascal Context datasets show that NSNPFormer achieves mIoU scores of 53.7 and 58.06, respectively, highlighting its effectiveness in semantic segmentation tasks.

Enhanced outdoor visual localization using Py-Net voting segmentation approach.

Camera relocalization determines the position and orientation of a camera in a 3D space. Althouh methods based on scene coordinate regression yield highly accurate results in indoor scenes, they exhibit poor performance in outdoor scenarios due to their large scale and increased complexity. A visual localization method, Py-Net, is therefore proposed herein. Py-Net is based on voting segmentation and comprises a main encoder containing Py-layer and two branch decoders. The Py-layer comprises pyramid convolution and 1 × 1 convolution kernels for feature extraction across multiple levels, with fewer parameters to enhance the model's ability to extract scene information. Coordinate attention was added at the end of the encoder for feature correction, which improved the model robustness to interference. To prevent the feature loss caused by repetitive structures and low-texture images in the scene, deep over-parameterized convolution modules were incorporated into the seg and vote decoders. Landmark segmentation and voting maps were used to establish the relation between images and landmarks in 3D space, reducing anomalies and achieving high precision with a small number of landmarks. The experimental results show that, in multiple outdoor scenes, Py-Net achieves lower distance and angle errors compared to existing methods. Additionally, compared to VS-Net, which also uses a voting segmentation structure, Py-Net reduces the number of parameters by 31.85% and decreases the model size from 236MB to 170 MB.

Learned liquid crystal microlens array for joint optimized deep optical architecture in identifying metameric materials.

Multispectral imaging holds great promise for the detection of metameric materials. However, traditional multispectral imaging systems are characterized by their large volume, complex structure, and high computational requirements, limiting their practical application. We propose a jointly optimized deep optical architecture that combines the liquid crystal (LC) microlens array (MLA) characteristics and a multi-level perceptual spectral reconstruction network (MLP-SRN). The core of the architecture is to integrate the physical properties of the LC-MLA into the MLP-SRN using point spread function (PSF) optical convolution kernels, decoupling the light-field characteristic information collected by the LC-MLA at different voltages. Experimental results demonstrate that the incorporation of the physical properties of the LC-MLA not only reduces the system size and computational complexity but demonstrates excellent performance in identifying a metameric material.

Concurrent Image Differentiation and Integration Processings Enabled By Polarization‐Multiplexed Metasurface

AbstractOptical computing and image processing performed by sensor front‐end metasurfaces is receiving increasing interest because of advantages such as significant reduction of latency time, energy consumption, and system complexity. Despite the rapid progress, concurrent processing, the most important feature of electronic computing, has not yet been well implemented in optical computing. Here, a metasurface‐based optical image processor that can perform optical differentiation and integration tasks simultaneously is proposed. This optical front‐end processor integrates two coherent transfer functions corresponding to differential and integral convolution kernels into a built‐in metasurface by polarization encoding, allowing concurrent processing of multiple all‐optical computational tasks. The simultaneous differentiation and integration operations on images for edge enhancement and denoising are demonstrated at multiple visible wavelengths. This concurrent processing architecture paves a promising pathway toward multifunctional and higher‐speed image processing for machine vision and biomedical imaging and shows the potential to expedite and potentially supplant certain digital neural network algorithms.

Improving Differential-Neural Cryptanalysis

Our first objective is to enhance the capabilities of differential-neural distinguishers by applying more deep-learning techniques, focusing on handling more rounds and improving accuracy. Inspired by the Inception Block in GoogLeNet, we adopted a design that uses multiple parallel convolutional layers with varying kernel sizes before the residual block to capture multi-dimensional information. Additionally, we expanded the convolutional kernels in the residual blocks, enlarging the network's receptive field. In the case of Speck32/64, our efforts yield accuracy improvements in rounds 6, 7, and 8, enabling the successful training of a 9-round differential-neural distinguisher. As for Simon32/64, we developed a differential-neural distinguisher capable of effectively handling 12 rounds while achieving noteworthy accuracy enhancements in rounds 9, 10, and 11. Additionally, we utilized neutral bits to ensure the required data distribution for launching a successful key recovery attack when using multiple-ciphertext pairs as input for the neural network. Meanwhile, we redefined the formula for time complexity based on the differences in prediction speeds of the distinguisher between a single-core CPU and a GPU. Combining these various advancements allows us to considerably reduce the time and data complexity of key recovery attacks on 13-round Speck32/64. Furthermore, we used knowledge distillation techniques to reduce the model size, accelerating the distinguisher's prediction speed and reducing the time complexity. In particular, we achieved a successful 14-round key recovery attack by exhaustively guessing a 1-round subkey. For Simon32/64, we accomplished a 17-round key recovery attack for the first time and reduced the time complexity of the 16-round key recovery attack.

Improved YOLOv8n for Lightweight Ship Detection

Automatic ship detection is a crucial task within the domain of maritime transportation management. With the progressive success of convolutional neural networks (CNNs), a number of advanced CNN models have been presented in order to detect ships. Although these detection models have achieved marked performance, several undesired results may occur under complex maritime conditions, such as missed detections, false positives, and low detection accuracy. Moreover, the existing detection models endure large number of parameters and heavy computation cost. To deal with these problems, we suggest a lightweight ship model of detection called DSSM–LightNet based upon the improved YOLOv8n. First, we introduce a lightweight Dual Convolutional (DualConv) into the model to lower both the number of parameters and the computational complexity. The principle is that DualConv combines two types of convolution kernels, 3x3 and 1x1, and utilizes group convolution techniques to effectively reduce computational costs while processing the same input feature map channels. Second, we propose a Slim-neck structure in the neck network, which introduces GSConv and VoVGSCSP modules to construct an efficient feature-fusion layer. This fusion strategy helps the model better capture the features of targets of different sizes. Meanwhile, a spatially enhanced attention module (SEAM) is leveraged to integrate with a Feature Pyramid Network (FPN) and the Slim-neck to achieve simple yet effective feature extraction, minimizing information loss during feature fusion. CIoU may not accurately reflect the relative positional relationship between bounding boxes in some complex scenarios. In contrast, MPDIoU can provide more accurate positional information in bounding-box regression by directly minimizing point distance and considering comprehensive loss. Therefore, we utilize the minimum point distance IoU (MPDIoU) rather than the Complete Intersection over Union (CIoU) Loss to further enhance the detection precision of the suggested model. Comprehensive tests carried out on the publicly accessible SeaShips dataset have demonstrated that our model greatly exceeds other algorithms in relation to their detection accuracy and efficiency, while reserving its lightweight nature.

Wind Turbine Bearing Failure Diagnosis Using Multi-Scale Feature Extraction and Residual Neural Networks with Block Attention

Wind turbine rolling bearings are crucial components for ensuring the reliability and stability of wind power systems. Their failure can lead to significant economic losses and equipment downtime. Therefore, the accurate diagnosis of bearing faults is of great importance. Although existing deep learning fault diagnosis methods have achieved certain results, they still face limitations such as inadequate feature extraction capabilities, insufficient generalization to complex working conditions, and ineffective multi-scale feature capture. To address these issues, this paper proposes an advanced fault diagnosis method named the two-stream feature fusion convolutional neural network (TSFFResNet-Net). Firstly, the proposed method combines the advantages of one-dimensional convolutional neural networks (1D-ResNet) and two-dimensional convolutional neural networks (2D-ResNet). It transforms one-dimensional vibration signals into two-dimensional images through the empirical wavelet transform (EWT) method. Then, parallel convolutional kernels in 1D-ResNet and 2D-ResNet are used to extract multi-scale features, respectively. Next, the Convolutional Block Attention Module (CBAM) is introduced to enhance the network’s ability to capture key features by focusing on important features in specific channels or spatial areas. After feature fusion, CBAM is introduced again to further enhance the effect of feature fusion, ensuring that the features extracted by different network branches can be effectively integrated, ultimately providing more accurate input features for the classification task of the fully connected layer. The experimental results demonstrate that the proposed method outperforms other traditional methods and advanced convolutional neural network models on different datasets. Compared with convolutional neural network models such as LeNet-5, AlexNet, and ResNet, the proposed method achieves a significantly higher accuracy on the test set, with a stable accuracy of over 99%. Compared with other models, it shows better generalization and stability, effectively improving the overall performance of rolling bearing vibration signal fault diagnosis. The method provides an effective solution for the intelligent fault diagnosis of wind turbine rolling bearings.

Few-shot bioacoustic event detection utilizing spectro-temporal receptive field

Bioacoustic event detection is a task of identifying specific animal sounds within a biological audio recordings. Specialists only can label temporal information of animal sounds since knowledge for biological sound is demanded for labeling and they can give only few annotations due to high labor intensity caused by large duration of biological recordings. Therefore, the task is set to few-shot learning scheme and method called prototypical network effectively deal with the few-shot learning scheme by learning representation space which represent each class for given few examples. In this work, we utilized spectro-temporal receptive field, which is inspired auditory cortex which responses actively to certain spectro-temporal modulation, as a convolutional layer kernel of prototypical network. Bioacoustic events retain plentiful spectro-temporal modulation that STRF is expected to capture animal sounds effectively. Also, STRF kernels are constructed to fixed shape rather than trained that a model utilizing STRF kernels would learn representation space with less parameter. We built a model called Two Branch STRFNet (TB-STRFNet), in which STRF branch captures spectro-temporal modulation by STRF kernels and the other branch captures detailed time-frequency information which can be smashed in STRF branch. TB-STRFNet outperformed other models showing that effectiveness of auditory system inspired method.

Fault Diagnosis of Induction Motors under Limited Data for Across Loading by Residual VGG-Based Siamese Network

This study proposes an improved few-shot learning model of the Siamese network residual Visual Geometry Group (VGG). This model combined with time–frequency domain transformation techniques effectively enhances the performance of across-load fault diagnosis for induction motors with limited data conditions. The proposed residual VGG-based Siamese network consists of two primary components: the feature extraction network, which is the residual VGG, and the merged similarity layer. First, the residual VGG architecture utilizes residual learning to boost learning efficiency and mitigate the degradation problem typically associated with deep neural networks. The employment of smaller convolutional kernels substantially reduces the number of model parameters, expedites model convergence, and curtails overfitting. Second, the merged similarity layer incorporates multiple distance metrics for similarity measurement to enhance classification performance. For cross-domain fault diagnosis in induction motors, we developed experimental models representing four common types of faults. We measured the vibration signals from both healthy and faulty models under varying loads. We then applied the proposed model to evaluate and compare its effectiveness in cross-domain fault diagnosis against conventional AI models. Experimental results indicate that when the imbalance ratio reached 20:1, the average accuracy of the proposed residual VGG-based Siamese network for fault diagnosis across different loads was 98%, closely matching the accuracy of balanced and sufficient datasets, and significantly surpassing the diagnostic performance of other models.

Phase unwrapping via fully exploiting global and local spatial dependencies

Phase unwrapping (PU) is the process of extracting the authentic phase image from its noisy wrapped measurements, playing a crucial role in scientific imaging techniques. PU requires solving a challenging non-linear ill-posed problem, particularly in the presence of noticeable noise. In recent years, deep learning has emerged as a promising approach for PU. Inspired by the success of convolutional neural networks (CNNs) in image restoration, many existing works trained CNNs for PU. However, due to the locality of convolutional kernels, CNNs are not efficient in capturing global spatial dependencies, a critical cue for PU. As an alternate, recent studies employed recurrent neural networks (RNNs) defined on handcrafted pixel paths. Nonetheless, a limited number of pre-defined pixel paths cannot fully exploit global spatial dependencies existing in complex phase structures. In this paper, we introduce a vision transformer (ViT) model that effectively captures both global and local spatial dependencies using a hierarchical structure with a multi-scale process. The proposed ViT model employs a series of global transformer blocks to capture global spatial dependencies at the roughest scale. The resulting global features are used to guide a set of local transformer blocks to analyze local spatial dependencies in a coarse-to-fine progressive manner for unwrapping. Extensive experiments show that, our proposed ViT model produces higher-quality unwrapped phases over existing CNN/RNN-based methods, while maintaining a lightweight nature.

An efficient CNN accelerator for pattern-compressed sparse neural networks on FPGA

Currently, the sparsity of weights and activations are mainly utilized to improve the energy efficiency and computational performance of CNN accelerators. However, the irregular sparsity of weights and activations can lead to problems such as mapping access conflict and imbalanced workload, which poses great challenges to the efficient computation of sparse networks. To address those problems, this paper proposes a CNN accelerator for pattern-compressed sparse neural networks on FPGA to achieve efficient inference acceleration. It mainly includes three parts: first, we propose a convolutional kernel grouping fusion method to achieve a balanced workload of weights. Second, we propose an activation hash mapping conflict reduction method to drastically reduce the mapping access conflict rate. Finally, we further propose a hierarchical mapping computing architecture to achieve efficient processing of mapping access conflicts and balanced adaptation of weight loads. To verify the effectiveness of the method proposed in this paper, our accelerator is designed and implemented on the Zynq UltraScale+ MPSoC ZCU102 evaluation board and evaluated by running VGG16 and ResNet50 networks. The experimental results show that the method proposed in this paper can reduce the mapping conflicts by more than 97 % and improve the utilization rate of processing element (PE) by 1.67×. Furthermore, the accelerator implemented in this paper can achieve 393.2 GOPs throughput, 1.40–3.25× computational performance, and 1.11–7.34× energy efficiency improvement.