Fig. 1. The training pipeline of the proposed ACE-STDN framework. Our method consists of two subnetworks to preprocess the infrared image and detect small targets respectively. The contrast enhancement subnetwork aids the small target detection subnetwork to achieve better performance，especially for dim targets.

Fig. 2. The adaptive contrast enhancement subnetwork for infrared images. This network consists of three main modules，where gray arrows denote convolution layers，and the green ones are deconvolution layers

Fig. 3. The structure of the transformer encoder block and TSConv block.

Fig. 4. The architecture of the detection subnetwork

Fig. 5. Two different frameworks：（a）the framework of YOLOv5；（b）our improved framework

Fig. 6. Infrared small-dim targets in the real world and their local intensity distribution：（a）simple background；（b）complex background

Fig. 7. The schematic diagram of measurements using a discrete bounding box and 2D Gaussian Distribution

Fig. 8. Illustration of detection results on ATDT

Fig. 9. Illustration of detection results on SIRST

Fig. 10. Illustration of detection results on a multiclass infrared dataset

Fig. 11. Illustration of detection results on a multiclass RGB dataset

Table 1. Ablation study on ATDT

Table 2. Comparison with generic detection method on ATDT

Table 3. Comparison with generic detection method on SIRST