To solve the problem of detection accuracy loss caused by reducing the number of parameters in lightweight target de-tection network,a target detection network MSPF-YOLOV3 was designed,which achieved the same information communication effect as high packet number through the channel shuffling structure with low packet number and low memory footprint.Shallow features with rich location information were fused with deep features to improve the detection accuracy of targets of different sizes.After experiments,the mAP of this network on PASCAL VOC07+12 dataset reaches 86.31％.In COCO2014 dataset,20 categories are extracted that are the same as PASCAL VOC07+12 dataset,and the mAP reaches 67.71％.The weight file size is 46.8 MB,showing a decrease of 198.2 MB compared to YoloV3.The detection speed reaches 44 FPS.The validity and the ability of real-time detection of the method are verified.
关键词
目标检测/轻量化/低分组数/低内存占用/通道混洗/特征融合/实时检测
Key words
target detection/lightweight/low number of clusters/low memory footprint/channel shuffling/feature fusion/real-time detection
维普
万方数据