Deep Learning Based Self-Attentive Feature Fusion Vehicle Detection Method
Aiming at the problems of insufficient model feature characterization ability,low detection accuracy,and overlapping leakage detection in the current vehicle target detection task,a self-attention feature fusion vehicle target detection model based on YOLOv8 improvement is proposed.First,a self-Attention Convolutional aggregation feature extraction module(AC-ReELAN)is proposed,which fuses ACMix and CBS using a stage-level method,combines multi-granularity feature information through residual links,improves the original network backbone network structure,and enhances the model feature information extraction and characterization ability;second,a DynamicConv by CBAM bi-dimensional attention mechanism(D_CBAM)is introduced in the Neck part by designing the adaptive convolutional value space,dynamically restricting the range of weights,and establishing a multi-channel weight assignment mechanism,which improves the filtering ability of the model while ensuring the model is streamlined,strengthens the attention to the target,and enhances the model's ability to accurately locate the model;finally,the MPDIoU bounding box regression loss function is introduced to optimize regression indexes,simplify the computation process,accelerate the model's convergence speed,and improve the model robustness.The experimental results show that the average detection progress of the improved model reaches 85.7％,which is 10.1 percentage points higher compared with Y OLOv8,and the detection speed reaches 51 frames per second,which meets the requirement of real-time detection.
NETL
NSTL
万方数据
