采用深度可分离卷积和改进特征融合的MSSD

单步多框检测器(Single Shot Multibox Detector,SSD)是一种优秀的目标检测模型,但是其对额外层的处理方式还需要进一步提升。因此,利用深度可分离卷积的思想设计新的深度可分离卷积模块改进模型中的额外层,采用紧邻特征图融合方法加强特征复用,综合设计了改进的目标检测模型(Modified SSD,MSSD)。该模型在VOC 2007和VOC 2012训练集上进行训练后,在VOC 2007测试集上进行测试。当输入尺寸为300×300时,它的平均精度均值(mean Average Precision,mAP)可达79.1%,相比原来的77.2%提高了1.9%,且检测速度可达55 f/s。同SSD的各类变体对比,MSSD的性能具有一定的优势,在速度和精度上取得了较好的平衡性。

MSSD with Depthwise Separable Convolution and Improved Feature Fusion

Single Shot Multibox Detector(SSD)is an excellent target detection model,but its processing method for outer frontal layer needs to be further improved.This article uses the idea of depthwise separable convolution to design a new depthwise separable convolution module to improve the extra layers in the model,and uses the adjacent features fusion method to strengthen the reuse of features,designs an improvement object detection model Modified SSD(MSSD).After the model is trained on the VOC 2007 and VOC 2012 training sets,it is tested on the VOC 2007 test set.When the input size is 300×300,the mean Average Precision(mAP)can reach 79.1%,which is an increase of 1.9 compared to the original 77.2%,the detection speed can reach 55 f/s.Compared with various variants of SSD,the performance of MSSD also has certain advantages,achieving a better balance between speed and accuracy.