Aiming at the problems of small target information loss caused by stride operation and large redundancy among multi-scale feature maps generated by serial structure in original additional feature extraction network (OAFEN) of SSD
a depthwise separable dilated convolution (DSDC) with small computation and large field of receptivity is proposed; then a parallel additional feature extraction network (PAFEN) with three independent subnetworks is designed by using five DSDCs. In upper subnetwork of PAFEN
two DSDCs are used to extract 19*19 and 3*3 feature maps. In intermediate subnetwork of PAFEN
one DSDC is used to extract 10*10 feature maps.In lower subnetwork of PAFEN
two DSDCs are used to extract 5*5 and 1*1 feature maps. The experimental results show that within the framework of SSD
PAFEN is superior to OAFEN in terms of mAP and detection time
and is suitable for ground small target detection tasks.