Fig. 1. Design idea of NCDN

Fig. 2. SSD framework based on feature fusion

Fig. 3. Convolutional architecture of MobileNet Depthwise/Pointwise

Fig. 4. Convolution channel compression

Fig. 5. Tango(左)与Mango(右)

Fig. 6. Local feature definition and label of SPEED

Fig. 7. Scale changes of target on SPEED

Fig. 8. Comparison effects of NCDN at different distances

Fig. 9. mAP comparison of model changes at different distances