Multi-scale Feature Fusion Enhanced Lightweight Detection

Abstract: With the advancement of automation technology, automated car wash systems have been widely utilized. However, Such technology consume large amounts of Water resources. Moreover, existing approaches face the challenge of standardized parameter settings, which make it difficult to adapt to variations in vehicle body structures and surface scratch features, thereby hindering effective stain removal while protecting the car body. Existing approaches focus on the investigation and development of car wash machines. They employ conventional image processing techniques such as threshold segmentation and mechanical water recycling optimization strategies to achieve improved cleaning efficiency and enhanced defect detection and cleaning precision. However, these traditional, statistics-based approaches face significant challenges in handling complex real-world conditions. They exhibit suboptimal detection accuracy, insufficient precision in dynamic water flow control, and often suffer from a lack of training datasets, while traditional algorithms struggle to balance precision with real-time constraints. These limitations underscore the need for advanced intelligent detection and control methods.To tackle these challenges, a vehicle defect dataset comprising 10,320 samples across 11 categories was constructed. Then, an optimized car wash model based on the RetNet architecture was developed, integrating a dynamic channel attention mechanism (DCAM) and a multi-scale feature fusion module to enhance the model's adaptability to complex environments. The model was subsequently trained and evaluated through simulation, demonstrating notable improvements in defect detection accuracy and cleaning efficiency. The contribution of this paper are as follows:1）A dedicated vehicle body defect dataset comprising 10,320 samples was constructed for model development and training, thereby alleviating the data scarcity issue in the intelligent cleaning domain.2）A lightweight detection algorithm based on the RetNet architecture was designed, incorporating a dynamic channel attention module and a multi-scale feature fusion mechanism to enhance both defect recognition accuracy and processing efficiency.3）A multi-round iterative optimization framework was implemented, integrating knowledge distillation and a hybrid loss function to systematically improve the model’s lightweight capability and small-target detection performance, ultimately achieving a synergistic optimization of cleaning strategies and resource utilization.Experimental results shows that the lightweight algorithm based on the improved RetNet achieves a Precision of 88%, a Recall of 87%, an Accuracy of 88%, and an F - Score of 87% on the self - built vehicle body defect dataset (10k - 12k images). Compared with the optimal baseline algorithm, it has improvements of 1.8%, 1.8%, 3.4%, and 2.6% respectively. Moreover, the generalization experiment results on the CIFAR - 10, STL - 10, ImageNet, and ObjectNet datasets demonstrate that the proposed scheme has good robustness. The experimental results demonstrate that our proposed RetNet optimization algorithm achieves robust performance across several evaluation metrics. On the CIFAR-10 dataset, the Precision, Accuracy, Recall, and F-score are 89%, 89%, 88%, and 89%, respectively. Similarly, on STL-10, these values are consistently 87%, while on ImageNet they reach 88%, 88%, 87%, and 88%. In the case of the ObjectNet dataset, the algorithm attains scores of 86%, 86%, 89%, and 87% for the four respective metrics. These findings indicate that the RetNet optimization algorithm can effectively solve the problem of detecting scratches on vehicle bodies.

Keywords: Intelligent Design, RetNet, Sustainability

DOI: 10.54941/ahfe1007037

Cite this paper:

​