HI-CMAIM: Hybrid Intelligence-Based Multi-Source Unstructured Chinese Map Annotation Interpretation Model

Remote Sensing, Journal Year: 2025, Volume and Issue: 17(2), P. 204 - 204

Published: Jan. 8, 2025

Map annotation interpretation is crucial for geographic information extraction and intelligent map analysis. This study addresses the challenges associated with interpreting Chinese annotations, specifically visual complexity data scarcity issues, by proposing a hybrid intelligence-based multi-source unstructured method (HI-CMAIM). Firstly, leveraging expert knowledge in an innovative way, we constructed high-quality knowledge-based dataset (EKMAD), which significantly enhanced diversity accuracy. Furthermore, improved detection model (CMA-DB) recognition (CMA-CRNN) were designed based on characteristics of both incorporating knowledge. A two-stage transfer learning strategy was employed to tackle issue limited training samples. Experimental results demonstrated superiority HI-CMAIM over existing algorithms. In task, CMA-DB achieved 8.54% improvement Hmean (from 87.73% 96.27%) compared DB algorithm. CMA-CRNN 15.54% accuracy 79.77% 95.31%) 4-fold reduction NED 0.1026 0.0242), confirming effectiveness advancement proposed method. research not only provides novel approach support but also fills gap high-quality, diverse datasets. It holds practical application value fields such as systems cartography, contributing interpretation.

Language: Английский

Vehicle and Pedestrian Detection Based on Improved YOLOv7-Tiny

Electronics, Journal Year: 2024, Volume and Issue: 13(20), P. 4010 - 4010

Published: Oct. 12, 2024

To improve the detection accuracy of vehicles and pedestrians in traffic scenes using object algorithms, this paper presents modifications, compression, deployment single-stage typical algorithm YOLOv7-tiny. In model improvement section: firstly, to address problem small missed detection, shallower feature layer information is incorporated into original fusion branch, forming a four-scale head; secondly, Multi-Stage Feature Fusion (MSFF) module proposed fully integrate shallow, middle, deep extract more comprehensive information. compression Layer-Adaptive Magnitude-based Pruning (LAMP) Torch-Pruning library are combined, setting different pruning rates for improved model. V7-tiny-P2-MSFF model, pruned by 45% LAMP, deployed on embedded platform NVIDIA Jetson AGX Xavier. Experimental results show that achieves 12.3% increase [email protected] compared with parameter volume, computation size reduced 76.74%, 7.57%, 70.94%, respectively. Moreover, inference speed single image quantized Xavier 9.5 ms.

Language: Английский