Non-GDANets: Sports small object detection of thermal images with Non-Glodal decoupled Attention

Abstract

Because thermal infrared sport targets have rich and complex semantic information, there is a high coupling between different types of features. In view of these limitations, we propose a Non-Glodal decoupled Attention, namely,local U-shaped attention decoupling network (LUANets), which aims to decompose the coupling relationship of different sport target features in thermal infrared images and establish effective spatial dependence between them. This method takes the captured multi-scale initial features according to different levels and inputs them into the local decoupling module with U-shaped attention structure to realize the decomposition of semantic details. At the same time, considering the correlation between different targets, in the process of feature decomposition, using prior knowledge as guiding information many times to establish effective spatial dependence. Secondly, we design a two-way cross-aggregation FPN module to cross-aggregate information flows in the front and back directions to achieve feature interaction while further reducing the coupling between different types of features. The evaluation results on data such as TIIs,SportFCs and FLIR show that the LUANets method we proposed has achieved the best detection performance, with mAP of 68.72%,59.51% and 65.29%, respectively.

LAGSwin: Local attention guided Swin-transformer for thermal infrared sports object detection

“…RGBsports [ 16 ]. This dataset contains 3000 RGB images, including 1874 footballs and 1126 crickets.…”

“…For example, Masuda T et al [15] proposed a motion video behavior detection method based on self-supervised feature learning and target detection, which introduced target detection into the process and realized the action detection of multiple people by tracking each person. Considering the high coupling between different features, Zhao J et al [16] designed a non-global attention mechanism: a local ushaped attention decoupling network. Jiang X et al [17] propose a new complementary transformer network (MCNet) for object detection in RGB and thermal infrared images, that is, introduce a transformer-based feature extraction module to efficiently extract hierarchical features of RGB and thermal images and attention-based feature interaction and serial multiscale dilated convolution (SDC)-based feature fusion module, the complementary interaction of low-level features and semantic fusion of deep features are realized.…”