424
J. Eur. Opt. Society-Rapid Publ. 22, 43( 2026)
Figure
7. Constraint unit for lightweight deployment: Balancing computational efficiency and feature representation capability of the YOLO-KMM model.
Table 1. Ablation experiment results of YOLO-KMM model. |
Detection algorithm |
|
Module |
|
|
Result |
|
|
( lr) 2-4( lr) 5-8 |
C2KD |
C3K2-MU |
ConvM |
mAP 50 |
mAP 50�95 |
P(%) |
R(%) |
|
BASE
+ C2KD
+ C2KD + C3K2-MU
YOLO-KMM
|
p p p |
p p |
p |
80.4
82.8
84.1
88.2
|
42.5
45.3
47.2
48.4
|
86.9
91.9
92.6
94.0
|
71.4
72.3
72.8
74.3
|
Figure 8. The following heatmap diagrams figures illustrate the YOLOv11 prediction results and other improve modules.
3 Experiments and results
3.1 Experimental environment
All experiments in this study were conducted on a GeForce RTX 2080 graphics card with 8 GB of video memory. The software environment was configured as follows: Windows 11 operating system, CUDA 12.4 acceleration library, Python 3.13.9 programming language, and PyTorch deep learning framework. The model training was optimized using the Stochastic Gradient Descent( SGD) optimizer, with the batch size set to 12, the initial learning rate configured as 0.01, and the learning rate decay coefficient set to 0.0005. The resolution of input experimental images was uniformly fixed at 640 640 pixels.
3.2 Evaluation metrics
To verify the performance of the proposed model, this study selects the following metrics for quantitative evaluation: Precision( P), Recall( R), mean average precision( mAP 50), comprehensive mean average precision