In order to improve the detection performance of DSOD for small objects, the RFB_a network module and Atrous convolution were introduced into DSOD. Firstly, the improved algorithm inputs the feature map generated by the second transition layer of the DSOD network into the RFB_a network module, and extracts the features with different receptive fields through the Atrous convolution with different sampling steps of the RFB_a network for small objects detection. Then, in order to improve the semantic information in the feature map, an Atrous convolution layer with rate size of 6 was added after the second transition layer without pooling. Finally, an IOG penalty term was added to the loss function to prevent the prediction box from overlapping when predicting objects of the same class, which may leads to detection missing after NMS processing. Experimental results show that, compared with the original DSOD algorithm, the improved algorithm can obtain higher detection accuracy and better detection ability of small objects, and meanwhile reduce the hardware requirements on the training network.
