CDF-YOLOv8: City Recognition System Based on Improved YOLOv8

Abstract

To address challenges in urban traffic management, especially detection under low exposure conditions and image quality degradation caused by weather factors, this paper proposes an urban detection algorithm based on the YOLOv8 model. Initially, The idea introduced the Chain-of-Thought Prompt Adaptive Enhancer (CPA-enhancer) to enhance image processing capabilities to cope with unknown image quality degradation. Secondly, the lightweight and efficient dynamic module DySample replaces the original upsampling module, boosting the model’s upsampling capability. Furthermore, YOLOv8’s Spatial Pyramid Pooling Fast (SPPF) was replaced with FocalModulation to enhance feature processing, particularly for small objects. Finally, experimental results show that compared to the original model, our enhanced algorithm achieved significant improvements in the precision, recall, and map50, with values of 70%, 53%, and 0.6056 respectively. These improvements represent increases of 4.48%, 8.16%, and 8.12%. Our enhanced algorithm surpasses both YOLOv8 and YOLOv10 in recognizing urban traffic imaging degradation.

Keywords

• Noise image recognition
• deep learning
• YOLOv8
• CPA-enhancer
• DySample
• FocalModulation