An improved mobilenetv3 feature extraction network
A feature extraction and network technology, applied in the backbone network field, can solve the problem of reducing model accuracy and achieve the effect of less calculation and better classification accuracy
Active Publication Date: 2022-04-29
UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF29 Cites 0 Cited by
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
[0006] However, the method of reducing the model or compressing feature information purely through efficient convolution and other methods will inevitably reduce the accuracy of the model, although the amount of parameters and calculations will improve the model.
Method used
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View moreImage
Smart Image Click on the blue labels to locate them in the text.
Smart ImageViewing Examples
Examples
Experimental program
Comparison scheme
Effect test
Embodiment Construction
[0028] The present invention will be further described below in conjunction with the accompanying drawings.
[0029] First of all, the network model based on convolutional neural network is in the process of feature extraction as follows: figure 1 shown. Input image X∈H×W×C in , where H, W and C in are the height and width of the input image and the number of feature channels, respectively, and the output feature map Y∈H×W×n can be obtained by processing n convolution filters with a size of k×k. Randomly select the feature map of a certain layer for visualization, such as figure 2 As shown, it can be found that there are many similar feature map pairs in the feature set. For above-mentioned characteristic, the present invention proposes Shadow Module module, as image 3 As shown, in this module, first use some convolution methods to generate a small number of ontology feature maps, and then use some cheaper calculation methods to obtain their shadow feature maps on these...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More PUM
Login to View More Abstract
The invention discloses an improved MobileNetV3 feature extraction network. This model has certain versatility in the field of CNN-based computer vision technology. The model with Resnet and Vgg as the core ignores the large redundancy and similarity of its feature set when the image is processed by the feature extraction network, and there are problems of high parameter amount and large amount of calculation. Aiming at the problems of redundancy and similarity, a shadow-bottleneck is proposed, that is, by using group convolution and improved channel shuffling to generate a small number of ontology features, and then using efficient operations to generate shadow features to ensure the richness of features and redundancy; for the lightweight problem, refer to the MobileNetV3 model structure and replace the bottleneck in the network with shadow-bottleneck to form a final improved lightweight feature extraction network model. The model can have a low amount of computation and parameters, and can obtain high classification accuracy.
Description
technical field [0001] The invention relates to the field of backbone networks used for feature extraction in deep learning. Background technique [0002] The construction of algorithm models such as object detection and semantic segmentation in the field of computer vision is inseparable from the support of CNN. At this stage, common feature extraction networks such as Resnet obtain higher classification accuracy by building a relatively large model. However, deeper The network will add more parameters and calculations to the network. Taking Resnet-101 as an example, the model parameters are about 46.5M, and the floating point calculations are 7.6B, which is not real-time and lightweight. [0003] With the development of technology and the evolution of requirements, people pay more and more attention to lightweight models. At present, common lightweight network models are roughly divided into: [0004] (1) MobileNet series: Proposed to use point-by-point convolution (PW) a...
Claims
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More Application Information
Patent Timeline
Login to View More Patent Type & Authority Patents(China)
IPC IPC(8): G06V10/44G06V10/764G06V10/82G06K9/62G06N3/04G06N3/08
CPCG06N3/08G06V10/44G06N3/045G06F18/241
Inventor 贾宇明唐昊贾海涛田浩琨王子彦王云邹新雷
Owner UNIV OF ELECTRONICS SCI & TECH OF CHINA