Share Email Print

Proceedings Paper

An improved YOLOv2 model with depth-wise separable convolutional layers for object detection
Author(s): Zhuo Han; Dongfei Wang; Aidong Men; Yun Zhou
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

Object detection is the basic research direction in the field of computer vision. It provides basic image information data for other advanced computer vision processing and analysis tasks. With the continuous breakthrough of deep machine learning technology, especially convolutional neural network model in the field of digital image processing shows a strong ability to extract image features. By choosing the depth separable convolution layer to replace the standard convolution layer used in the traditional model, the number of parameters of CNN network model is compressed. Depth Separable Convolution Layer (DSCL) decomposes the standard convolution layer factor into depth convolution layer and point convolution layer, and extracts and merges image features in two steps to reduce the number of parameters. By introducing a depth-separable convolution layer instead of a standard convolution layer, the number of parameters of the model convolution layer is reduced by 78.1%. We choose image feature pyramid network to fuse the image features extracted from each layer of CNN network, so that the target detection model can use matching image fusion features for different size and shape of the target to be detected. The average detection precision on the PASCAL VOC dataset increased to 77.5%.

Paper Details

Date Published: 6 May 2019
PDF: 8 pages
Proc. SPIE 11069, Tenth International Conference on Graphics and Image Processing (ICGIP 2018), 110693V (6 May 2019); doi: 10.1117/12.2524181
Show Author Affiliations
Zhuo Han, Beijing Univ. of Posts and Telecommunications (China)
Dongfei Wang, Academy of Broadcasting Science (China)
Aidong Men, Beijing Univ. of Posts and Telecommunications (China)
Yun Zhou, Academy of Broadcasting Science (China)

Published in SPIE Proceedings Vol. 11069:
Tenth International Conference on Graphics and Image Processing (ICGIP 2018)
Chunming Li; Hui Yu; Zhigeng Pan; Yifei Pu, Editor(s)

© SPIE. Terms of Use
Back to Top
Sign in to read the full article
Create a free SPIE account to get access to
premium articles and original research
Forgot your username?