Deep learning networks always compromise between speed and accuracy for their in-depth feature extraction. In this paper, we present a modified single shot multibox detector (SSD) model to achieve high speed while maintaining satisfactory accuracy for target detection. Firstly, the operational parameters are reduced by deleting the convolution layers and reducing the channels within. Thus, the parameters are reduced by 50% with a permissible precision loss, and the detection speed of the model is significantly improved. Secondly, a light multiple dilated convolution (LMDC) operator is introduced to compensate for the precision loss. The LMDC functions as a filter to extract global and semantic information from the feature map, thereby making feature information completer and more accurate. Moreover, to reduce the computation quantity and increase the computation efficiency of the network, the feature extraction and fusion of the convolution layer are separated. It transforms the complex multiplication into addition among the parameters. Finally, the LMDC-SSD is evaluated on 3 datasets for 300 × 300-sized inputs. It yields 98.99% mean average precision (mAP) and 85 frames per second for the apple datasets. The speed and accuracy are improved by 44% and 8.1%, respectively, compared to the original model. The speed and accuracy are improved by 0.99% and 65.71%, respectively, for the bicycle and person datasets.The speed and accuracy are improved by 0.26% and 112.9%, respectively, for the vehicle datasets. The experimental results have shown that the proposed LMDC-SSD is rather promising for detection with high detection speed and accuracy performance.

深度学习网络始终会在速度和准确性之间进行折衷，以进行深度特征提取。在本文中，我们提出了一种改进的单发多盒检测器（SSD）模型，可以在保持令人满意的目标检测精度的同时实现高速。首先，通过删除卷积层并减少其中的信道来减少操作参数。因此，将参数减少了50％，并允许了精度损失，并且显着提高了模型的检测速度。其次，引入了一个轻量的多重扩张卷积（LMDC）运算符来补偿精度损失。LMDC用作从特征图提取全局信息和语义信息的过滤器，从而使特征信息更完整，更准确。而且，为了减少网络的计算量，提高网络的计算效率，将卷积层的特征提取与融合分离开来。它将复数乘法转换为参数之间的加法。最后，在3个数据集中针对300×300大小的输入评估LMDC-SSD。苹果数据集的平均平均精度（mAP）为98.99％，每秒85帧。与原始模型相比，速度和准确性分别提高了44％和8.1％。自行车和人的数据集的速度和准确性分别提高了0.99％和65.71％，车辆的数据集的速度和准确性分别提高了0.26％和112.9％。