To address the issues of inconvenient operation,low efficiency,and difficulty in managing recognition data during the manual identification of substation layout drawings,this paper proposes a morphology-based large-size drawing segmentation method and a text-image multimodal fusion drawing symbol detection method.Combined with post-processing methods for symbol detection,this forms a detectable and adaptable approach to large-size layout drawing symbol detection that can be generalized to other fields.The text-image multimodal fusion drawing symbol detection model is improved upon the open-set object detection model YOLO-World,by introducing the CTCM,SOFEM,and CJFFM.These enhancements significantly improve the model's performance in symbol recognition.Using the proposed methods,the detection of symbols in actual high-speed railway traction substation general layout drawings dataset is achieved.Compared to the original model,the proposed improved model,while maintaining a similar level of complexity,reaches an average precision of 97.5% for symbol recognition,with mAP@50:95 and mAP@90 increasing by 1.1% and 3.0%,respectively.
In recent years,the technologies of LLMs have been rapidly developed,with their applications across various industries experiencing vigorous growth.From natural language processing to intelligent recommendations,and from information retrieval to automated writing,LLMs are becoming indispensable tools in many fields.However,with the diversification of application scena-rios and the increase in demands,the efficiency of LLM inference is becoming increasingly prominent.In practical applications,ra-pid and accurate inference capabilities are crucial for responding to user queries,handling large-scale data,and making real-time decisions.To address this challenge,academia has undertaken extensive research and exploration to enhance the inference efficiency of LLMs.This paper comprehensively surveys the literature on efficient LLM inference in practical application scenarios.Firstly,it introduces the principles of LLMs and analyzes how to improve LLM inference efficiency in practical application scenarios.Secondly,it proposes a taxonomy tailored for real-world applications,which consists of three main levels:algorithm optimization,parameter optimization,and system optimization.This survey summarizes and categorizes related work about LLMs.Finally,it discusses potential future research directions.
Event cameras based on neuromorphic spike signals can provide information about illumination changes,compensating for the performance degradation of traditional RGB cameras in object detection under adverse environments.However,existing methods fusing event cameras with conventional cameras suffer from large model parameters and non-end-to-end training approaches,which restrict the effectiveness of modality fusion.To address this,this paper proposes a lightweight end-to-end object detection framework that integrates event and RGB information through multi-granularity fusion of multi-scale features across different network levels.By implementing lightweight fusion modules with reparameterized convolutions and enabling end-to-end training,the proposed framework enhances the model’s capability to extract complementary information from both modalities,overcoming challenging conditions in autonomous driving.Evaluated on the large-scale PKU-SOD dataset containing vehicular visual data under low-light,high-speed motion blur,and normal illumination scenarios,the proposed method significantly reduces model parameters compared to state-of-the-art multimodal approaches while improving detection accuracy and inference speed,demonstrating superior performance over existing methods.
In social media,people’s comments usually describe a certain sentiment region in the corresponding image,and there is correspondence information between image and text.Most previous multimodal sentiment analysis methods only explore the interactions between images and text from a single perspective,capturing the correspondence between image regions and text words,leading to results that are not optimal.In addition,data on social media is strongly personal and subjective,and the sentiment in the data is multidimensional and complex,which leads to the emergence of data with weak image and text sentiment consistency.To address the above two problems,a multimodal sentiment analysis model with interactive fusion of two perspectives under cross-modal inconsistency perception is proposed.On the one hand,cross-modal interaction of graphic and textual features from both global and local perspectives provides a more comprehensive and accurate sentiment analysis,which improves the perfor-mance and application of the model.On the other hand,the inconsistency scores of the graphical features are calculated to representthe degree of graphical inconsistency,as a way to dynamically regulate the weights of the unimodal and multimodal representations in the final sentiment features,thus improving the robustness of the model.Extensive experiments are conducted on two public datasets,MVSA-Single and MVSA-Multiple,and the results demonstrate the validity and superiority of the proposed multimodal sentiment analysis model compared to the existing baseline models,with F1 values increasing by 0.59 persentage points and 0.39 persentage points,respectively.
Screen-shooting resilient watermarking algorithms are of significant importance in fields such as copyright protection and traceability.Existing screen-shooting resilient watermarking algorithms mostly focus on natural images,neglecting research on document images.Document carriers inherently contain less redundant information,making it challenging to balance robustness and imperceptibility of the watermark.To address this issue,a screen-shooting resilient watermarking method for document image based on attention mechanism is proposed.To enhance the imperceptibility of the watermark,an attention feature fusion module is introduced in the encoder network to adaptively aggregate shallow and deep features.To improve the robustness of the algorithm for extraction,an adaptive channel-spatial attention module is designed in the decoder network to emphasize features that are particularly important in both channel and spatial dimensions.Additionally,a Moiré distortion layer is designed during screen-shoo-ting noise simulation to enhance the algorithm’s robustness against real Moiré distortions.Experimental results demonstrate that the proposed method achieves an average PSNR of 33.4 dB,SSIM of 0.988 5,RMSE of 5.48,and an average extraction accuracy of 99.49% in various screen-shooting scenarios.In terms of image quality and watermark robustness,the proposed method outperforms existing similar methods.
