In the huge automotive service aftermarket, the efficient and timely supply of maintenance parts has always attracted continuous concern from automotive service enterprises and end users,which is limited by the regional collaborative efficiency and information interaction among many suppliers over the automotive parts supply networks. However, the supply networks consist of enterprises with different manufacturing capabilities, and are filled with multisource, massive, heterogeneous information that contains multiple entities and overlapping triplet relation, leading to difficulties in achieving uniform representation and adaptive understanding of information. Entity-relation extraction is essential for unified information representation.In this paper, we devise an entity relationship extraction(ERE) method based on heterogeneous graph neural networks and entity feature fusion, which treats entities and relation as nodes in a graph, and iteratively integrates node representation to identify the most suitable node features for ERE tasks. The method introduces an innovative mechanism: firstly, we extract the subject entities and fuse their features into node representations using an attention mechanism; and then, the relations and object entities are jointly extracted to achieving end-to-end triplet extraction. Experiments are conducted using parts supply chain data from partners. The results validates the effectiveness of the method and obtain outstanding performance in the automotive parts supply chain(APSC) networks.

To address the poor adaptability of traditional dynamic hand gesture recognition (HGR) methods in handling inter-subject variations and their limited capacity to capture diverse hand shapes and motion patterns across users, this paper proposes a feature-enhanced multimodal fusion network (MGD-Net). The network adopts a feature-level fusion strategy to fully leverage the complementary advantages of depth images and 3D hand skeleton data, thereby improving the model’s ability to represent gestures under individual differences. Specifically, MGD-Net takes feature-enhanced grayscale depth images as input to a convolutional neural network (CNN) for spatial feature extraction, employs a bidirectional gated recurrent unit (BiGRU) to model temporal dependencies, and introduces a DeepGRU(D-GRU) module to effectively capture the dynamic variations of Three-Dimensional(3D) skeletal features. This architecture enables efficient multimodal feature fusion and precise gesture recognition. Experimental evaluations on two public datasets, DHG-14/28 and SHREC-2017, demonstrate that MGD-Net achieves competitive accuracy rates of 91.24% and 89.88% for 14 and 28 gesture classes on DHG-14/28, and 94.28% and 90.38% on SHREC-2017, respectively, showcasing its robust performance and generalization capability, particularly in scenarios involving significant user-specific morphological variations. This work contributes insights to the field of multimodal perception, with implications for the upcoming 2025 International Conference on Artificial Intelligence-Powered Image Processing and Computer Vision Technology (AIPCVT2025).

Effectively capturing the multi-scale, multi-channel, and boundary features of target images is beneficial for improving the accuracy of salient object detection. A multichannel feature enhancement module (MFEM) is constructed by concatenating two feature maps along the channel, performing convolution activation, dot product addition, and convolution operations. An edge feature enhancement module (EFEM) is constructed by performing an inverse attention mechanism and Gaussian-Laplace transform on the predicted map, channel-wise concatenation, and convolution operations with the feature submap. Multi-scale features are extracted using a SwinTransformer encoder and enhanced using the MFEM. The enhanced features are then fused using a cross-attention mechanism for multi-level features. The fused feature map, the high-frequency feature map obtained by Gaussian-Laplace transforming the predicted map, and the external edge features are input into the EFEM to obtain the salient object detection results. This results in a novel salient object detection model with multi-scale, multi-channel, and boundary feature enhancement: MfENet. Experimental results show that the Mean Absolute Error (MAE), S-measure and F-measure accuracies of our method on the HKU-IS dataset are 0.018,0.924 and 0.959 respectively. These results demonstrate the effectiveness of our method. Furthermore, this work highlights the importance of multi-scale and boundary awareness in enhancing salient object detection, and its contribution to the technical field relevant to the theme of this conference.

Efficient naming of inspection photos of transmission line towers is vital in the maintenance of power grid equipment. Current inspection photo naming methods are mainly manual, which is neither rapid nor effective. Research on inspection photo naming is limited due to a shortage of inspection image datasets and low image resolution. Hence, we gathered inspection photos of real tangent towers using drones and created an inspection photo dataset TTower-345 for automatic naming model training purposes. We proposed an automatic naming model, IELC (improved EfficientNet network and LBP classification model), based on this dataset. IELC comprises a dual-branch structure that integrates a jointly improved EfficientNet model and an local binary patterns (LBP) classification model. Experimental results indicated that the proposed dataset contains more diverse inspection image features, which in turn helped the model learn more features. In our experiments, our proposed automatic naming method achieved a classification accuracy of over 95% and demonstrated reliability by exhibiting good generalisability in practical scenarios.

The proposed method for joint entity and relation extraction integrates the tasks of entity extraction and relation classification by sharing the encoding layer. However, the method faces challenges due to incongruities in the contextual information captured by these subtasks, resulting in potential feature conflicts and adverse effects on model performance. To address this, we introduced a novel joint entity and relation extraction method that incorporates multi-module feature information enhancement (MFIE) ( https://github.com/liyao345496280/Relation-extraction ). We employ a relation awareness enhancement module for the entity extraction task, which directs the model’s focus towards extracting entities closely related to potential relations using a potential relation extraction module and an attention mechanism. For the relation extraction task, we implement an entity information enhancement module that uses entity extraction results to augment the original feature information through a gating mechanism, thereby enhancing relation classification performance. Experiments on the NYT and WebNLG datasets demonstrate that our method performs well. Compared to the state-of-the-art method, the F1 score on the NYT dataset improved by 0.7

Aiming at the complexity and large parameter size of you only look once version four (YOLO v4) object detection network, which cannot meet the requirements of lightweight deployment and real-time computation on mobile devices, a lightweight object detection network called Ghost-YOLONet is proposed. Firstly, the backbone network of YOLO v4 is replaced by the lightweight network GhostNet, and the decoupled fully-connected attention mechanism is integrated into the Ghost module to better capture global information. Then, the parallel structure of the spatial pyramid pooling module in YOLO v4 is changed to a serial structure to improve the model's execution efficiency. Comparative experimental results on the PASCAL VOC2007 and VOC2012 datasets show that compared with the YOLO v4 model, Ghost-YOLONet reduces the parameter size by 81.1%, the model volume by 81.5%, and achieves mAP@0.5 of 81.4%. Moreover, the FPS is improved, meeting the requirements of real-time detection tasks.

Efficient Convolution Operator (ECO) algorithms have achieved impressive performances in visual tracking. However, its feature extraction network of ECO is unconducive for capturing the correlation features of occluded and blurred targets between long-range complex scene frames. More so, its fixed weight fusion strategy does not use the complementary properties of deep and shallow features. In this paper, we propose a new target tracking method, namely ECO++, using deep feature adaptive fusion in a complex scene, in the following two aspects: First, we constructed a new temporal convolution mode and used it to replace the underlying convolution layer in Conformer network to obtain an improved Conformer network. Second, we adaptively fuse the deep features, which output through the improved Conformer network, by combining the Peak to Sidelobe Ratio (PSR), frame smoothness scores and adaptive adjustment weight. Extensive experiments on the OTB-2013, OTB-2015, UAV123, and VOT2019 benchmarks demonstrate that the proposed approach outperforms the state-of-the-art algorithms in tracking accuracy and robustness in complex scenes with occluded, blurred, and fast-moving targets.

Tracking targets accurately and robustly in visually complex environments poses a formidable challenge. To address this, capturing a resilient appearance representation is essential while augmenting the model's ability to generalize and cope with diverse challenges such as object deformation, variations in illumination, changes in scale, and motion blur. This paper presents a method for sturdy tracking in intricate scenarios, employing the efficient convolution operator (ECO) tracker. Our approach incorporates the 2 key concepts: a) extracting profound features via the Conformer network by increasing the number of underlying channels, and b) flexibly adjusting the fusion weight for shallow and profound features based on factors such as the peak-to-sidelobe ratio and the joint score of adjacent frame trajectory smoothness. This approach enhances the model's adaptability and generalization in intricate environments by capitalizing on the complementary aspects of deeper-layer and shallow-layer features. Experimental outcomes validate the algorithm's efficacy in tackling varied challenges related to target tracking in complex environments, ensuring robust tracking with consistently high accuracy.

In the task of constructing complex urban road scene semantic segmentation, continuous downsampling operation in the Context Path module of the Bisenet network may cause the loss of some pixel feature information. To solve this problem, a supervised block that can be inserted into the Context Path module was proposed. It consisted of 3*3 convolution, batch normalization and ReLU linear correction and 1*1 convolution. The downsampling operation of each layer was inserted into a supervised block to guide the training process, so that the current layer can obtain feature information from the previous layer and retain rich context information. In this way, feature representation was enhanced to improve semantic segmentation accuracy. The experimental results showed that the method in this paper had better semantic segmentation results on the Cityscapes and CamVid urban road datasets, and its MIOU evaluation index had increased by 2.3% and 3.8%, respectively.

针对YOLOv5网络模型的金字塔结构不能有效捕捉目标的跨尺度特征信息和存在梯度消失现象以及易导致目标检测精确率不高的问题,利用FPT(feature pyramid transformer)的注意力机制有效提取目标的跨尺度特征信息,把原YOLOv5网络模型中的FPN(feature pyramid network)和PAN(path aggregation network)结构替换为FPT,在FPT结构的两端加入跳跃连接(skip connection)并引入新的Mish激活函数,从而提出一种改进金字塔和跳跃连接的YOLOv5目标检测网络模型YOLO FS.在PASCAL VOC和MS COCO数据集上的对比实验结果表明,基于YOLO FS网络的目标检测在平均检测准确率、召回率和F1值上均有明显提升.

There are still challenges in the field of video understanding today, especially how to use natural language to describe the visual content in videos. Existing video encoder-decoder models struggle to extract deep semantic information and effectively understand the complex contextual semantics in a video sequence. Furthermore, different visual elements in the video contribute differently to the generation of video text descriptions. In this paper, we propose a video description method that fuses instance-aware temporal features. We extract local features of instances on the temporal sequence to enhance perception of temporal instances. We also employ spatial attention to perform weighted fusion of temporal features. Finally, we use bidirectional long short-term memory networks to encode the contextual semantic information of the video sequence, thereby helping to generate higher quality descriptive text. Experimental results on two public datasets demonstrate that our method achieves good performance on various evaluation metrics.

The ability of the separator to capture the context-detailed features of speech signals and the number of parameters directly affect the accuracy and efficiency of speech separation in time-domain speech separation network (TasNet). This paper combines lightweight external attention with convolution and extends external attention to channel dimension; while satisfying the fine-grained extraction and modeling of spatial-channel correlation, it maintains small parameters and computation. Convolutional position coding is also used to integrate the contextual relationship and relative position information of speech features better. The above module then applies as a separator in the encoder-decoder structure based on TasNet, and a new convolution-augment external attention model for time-domain speech separation is proposed: ExConNet. The comparative experimental results show that ExConNet achieves considerable accuracy of speech separation, while its model parameters and calculation amount are significantly reduced, which can better meet the need for efficiency of speech separation.

Because of high integration and narrow gap between solder joints in PCBA（printed circuit board assembly） of smart phone, there is some noise in the image of PCBA inevitably and the edge of circular solder joint is obviously bonded and broken. That easily leads to the low accuracy of circular solder joint contour extraction. Therefore, an adaptive extraction algorithm of PCBA circular solder joint contour is proposed. Firstly,the guided filtering algorithm is used to effectively suppress the noise and capture the edge gradient information of circular solder joint. Secondly, the solder joint region is segmented according to the adaptive OTSU segmentation algorithm, and the contour information of circular solder joint is extracted by 8-neighborhood boundary tracking algorithm and Canny operator. Finally, the contour information is filtered by using the contour geometric characteristics of circular solder joint and the edge sliding window mechanism to obtain a finer contour of circular solder joint. The experimental results on 90 PCBA image data sets collected by the experimental equipment and three typical methods show that the accuracy, recall and F-score value of the proposed algorithm are 96.96%, 97.88% and 97.42%, respectively, and the running time is 310 ms. The contour of PCBA circular solder joint can be extracted quickly and effectively.

In this paper, a solution is proposed to address the low accuracy in facial expression recognition that results from the 1×1 convolution dimensionality reduction of the Bottleneck in ResNet50. To do so, the authors introduce the Ghost module and depth separable convolution to replace the 1×1 and 3×3 convolutions in the Bottleneck, respectively,in order to preserve more of the original feature information and improve the feature extraction ability of the trunk branch. The Mish activation function is also used to replace the ReLU activation function in the Bottleneck, further enhancing the accuracy of facial expression recognition. To further improve the ability of the model to express important information, the authors also introduce an asymmetric residual attention block（ARABlock） between the improved Bottlenecks. The proposed method, which is referred to as the ghost asymmetric residual attention network（GARAN） model, shows high recognition accuracy on the FER2013 and CK+ facial expression datasets based on comparative experimental results.

In the current vehicle re-identification task based on CNN method, due to the loss of detail information of image caused by convolution and downsampling operation, the ability to distinguish similar vehicles is affected. In this paper, a new vehicle Re-ID method ViT-ReID using visual Transformer was proposed. In order to extract robust discriminative features to enhance the ability of the model to distinguish similar vehicles, a SIE module based on ViT backbone network was designed to effectively reduce the deviation of learning features caused by camera perspective changes. At the same time, in order to improve the performance of the model, the latest VPT method for fine-tuning of large-scale Transformers was introduced. Finally, the ViT-ReID model proposed in this paper was compared with other methods in the VeRi-776 and VehicleID datasets. The experimental results showed that the proposed method can effectively extract robust discriminative features in vehicle re-identification and distinguish similar vehicles with high accuracy.

Discriminative correlation filter (DCF) tracking algorithms are commonly used for visual tracking. However, we observed that different spatio-temporal targets exhibit varied visual appearances, and most DCF-based trackers neglect to exploit this spatio-temporal information during the tracking process. To address the above-mentioned issues, we propose a three-way adaptive spatio-temporal correlation filtering tracker, named ASCF, that makes fuller use of the spatio-temporal information during tracking. To be specific, we extract rich local and global visual features based on the Conformer network, establish three correlation filters at different spatio-temporal locations during the tracking process, and the three correlation filters independently track the target. Then, to adaptively select the correlation filter to achieve target tracking, we employ the average peak-to-correlation energy (APCE) and the peak-to-sidelobe ratio (PSR) to measure the reliability of the tracking results. In addition, we propose an adaptive model update strategy that adjusts the update frequency of the three correlation filters in different ways to avoid model drift due to the introduction of similar objects or background noise. Extensive experimental results on five benchmarks demonstrate that our algorithm achieves excellent performance compared to state-of-the-art trackers.

为提升布谷鸟算法对高维问题空间的搜索能力及算法的稳定性,提出了一种参数动态更新的布谷鸟搜索算法.该算法首先选用柯西随机数,动态更新迭代学习的步长因子,并结合Lévy飞行策略生成新解;在新解生成之后,提出了 一种正态扰动策略生成干扰解以增加解的搜索空间;对迭代中的新解和干扰解,采用模拟退火算法得到优势解,从而避免算法陷入局部最优,提升了算法的容差性;并提出一种轮盘赌选择和双向随机搜索策略强化迭代中优势解的学习.实验结果表明改进算法拥有较高的准确性和稳定性.

To solve the problem that the global features extracted by the ResNet-50 network has insufficient recognition capability in similar vehicle re-identification task, a new Re-ID method combining metric learning is proposed. Firstly, the fine-grained features of vehicles are extracted by using triplet constraints, and then combined with the global features extracted by the backbone network as vehicle features. Secondly, the similarity of different vehicle features is judged and ranked by Euclidean distance, so as to obtain more accurate results. Finally, a comparative experiment is conducted on the VeRi-776 dataset for different network models. The results show that our method has high recognition accuracy in Re-ID tasks. Compared with ResNet-50, the mean average accuracy (mAP) is improved by 2.30 %, rank-l increased by 2.31 %, and the rank-5 increased by 2.05 %. It is verified that this model can effectively improve the recognition accuracy in vehicle Re-ID.

为提高印刷电路板装配(PCBA)中目标区域检测的准确性和实时性,提出一种高斯金字塔变换与新粒子群优化算法结合的PCBA模板匹配算法.采用倒Sigmod函数调整粒子群迭代的惯性权值;分别构建个体和群体的自适应学习因子模型;提出粒子是否陷入局部解的自适应判据并对其采用随机动量因子进行调整,从而提出一种新的粒子群优化算法.分别对待匹配图像和模板图像进行4层高斯金字塔变换,采用新粒子群优化算法搜索待匹配顶层子图的粗匹配区域,该区域经高斯金字塔反变换后生成的邻域范围与对应的模板子图进行遍历匹配,在最底层得到最终匹配结果.对比实验结果表明,所提方法在PCBA模板匹配应用中具有准确性和实时性.

针对传统Canny算子不能有效滤除图像在解码处理和传输过程产生的椒盐噪声、无法保留边缘细节的问题,提出强噪声下Canny算子图像边缘检测算法.依据椒盐噪声的极值性、灰度差值性,将像元点划分为噪声点、疑似噪声点;根据分类之后的像元点自适应地改变滤波器窗口的大小和权值,在降低噪声影响的同时能较好地保留图像细节.引入8个方向模板的Sobel算子计算梯度幅值以提高滤波后的边缘定位效果.使用迭代自适应阈值算法与Otsu算法选择最佳阈值,实现阈值自适应设定,提高边缘连接效果.实验结果表明:图像去噪后的结构相似度为0.949,峰值信噪比相较于传统算法提升了10.97 dB.边缘评价指标提高27.2％,F1值提高了34.6％.该算法能有效去除椒盐噪声,具有更好的边缘细节保护能力.