An improved SSD target detection method based on deep separable convolution

This is the authors' abstract. We don't add key points for this paper.

Nowadays, the demand for target detection is increasing, in many scenarios, target detection model must run on mobile devices or embedded devices. Due to the large calculation of target detection model based on convolutional neural network and the limitation of hardware computing resources, it cannot meet the requirements of accuracy and speed at the same time. In order to solve this problem, this paper proposes an improved Single Shot Multibox Detector(SSD) method, the model is optimized by using depth separable convolution and using depth-wise convolution to complete down-sampling separately. The experimental results show that improved SSD algorithm has the least amount of parameters of the entire model compare with SSD and Moblienet V1-SSD. The mean average precision (mAP) on the test sets VOC2007 and VOC2012 reaches 73.14% and 66.6% respectively, which is higher than Moblienet V1-SSD and has small changes compared to SSD algorithm. Improved SSD algorithm improves the speed of the model and ensures the accuracy of the model.