Multi-Task Feature Decoupling Network with clear division of labor for vehicle component detection

Abstract

To enhance the precision and stability of vehicle component detection for intelligent claims assessment and intelligent vehicle detection, it is necessary to address the problem of missing detection and false detection in complex scenes caused by feature similarity between classes and insufficient feature extraction. In this paper, we propose a decoupling network architecture, named the Multi-Task Feature Decoupling Network (MTFDN), which is based on a two-stage multitasking algorithm. Three improved modules have been designed and integrated into the architecture for this purpose. First, we design an improved Parallel Dynamic Attention Module (PDAM) to facilitate a more scientific and logical acquisition of features and localization of regions of interest. Then, we introduce an improved Bidirectional Fusion Feature Pyramid Network (BF-FPN) module in the segmentation task, followed by the design of a Bidirectional Hybrid Dilated Fusion Feature Pyramid Network (BHDF-FPN) module in the detection task. The BF-FPN and BHDF-FPN modules are used to meet the different requirements of different tasks for features after decoupling. The experimental results demonstrate that, in comparison with the baseline model, MTFDN exhibits superior performance in terms of detection and segmentation on two vehicle-related datasets, namely the Vehicle Scene Dataset (VSD) and the Vehicle Component Dataset (VCD). Moreover, MTFDN demonstrates robust generalization capabilities when integrated with other two-stage algorithms, significantly enhancing the detection and segmentation performance of the algorithm itself, thereby highlighting its superiority in this domain.