Fig. 1. Detection flow of YOLO v3

Fig. 2. Dilated convolutions. (a) rrate=1; (b) rrate=2; (c) rrate=3

Fig. 3. Backbone network and FPN architecture of the improved YOLO v3

Fig. 4. Structure of dilated convolution residuals. (a) Dilated convolution bottleneck; (b) dilated convolution bottleneck with 1×1 Conv projection

Fig. 5. Two planes of occlusion

Fig. 6. Flow chart of the optimized NMS processing

Fig. 7. Loss curve

Fig. 8. Detecting results of multi-scale small targets by different methods

Fig. 9. Contrast experiments of aircraft detection with different occlusion proportions. (a)(b) Occlusion is close to 20%; (c)(d) occlusion is close to 60%; (e)(f) obvious color characteristics, occlusion is close to 60%; (g)(h) occlusion is close to 80%