Design method of human-shaped upper body monitoring network structure

Abstract

The invention provides a human-shaped upper body monitoring network structure design method, which comprises the following steps: S1, initial setting: setting an input size as m x m, wherein m > 24; when the stride is 1, the corresponding calculation formula is n-2, when the stride is 2, the corresponding calculation formula is (n-2) / 2, and n is the input size of each layer; as the final output is Nx3x3, and N is the number of channels, performing 3 * 3 convolution to finally regress the score and regression position coordinates of the face candidate box; s2, setting strides to be 2, reversely deducing the input size according to the calculation formula in S1, and obtaining the input size, form and calculation amount of each layer of the network; and S3, according to the network designed in the S2, carrying out first-stage training on the upper half body of the human shape.

Description

technical field [0001] The invention relates to the field of intelligent identification, in particular to a design method of a humanoid upper body monitoring network structure. Background technique [0002] With the continuous development of science and technology, especially the development of computer vision technology, intelligent recognition technology is widely used in various fields such as information security and electronic authentication, and the image feature extraction method has good recognition performance. [0003] Additionally, commonly used terms in the prior art include: [0004] Upper body detection: that is, to detect the middle position from the head to the elbow to the shoulder of the human body. [0005] Cascaded convolutional network: It is to design an N-order network to process tasks from coarse to fine. [0006] Recall rate: face as an example, the number of successfully detected face frames in the picture accounts for the percentage of the origin...

AI Technical Summary

Problems solved by technology

Although the speed of ONet is slow, but because of the high-probability frame has been obtained through the first two networks, there are fewer images input into ONet, and then ONet outputs accurate frame and key point information

[0010] There are also some problems in the existing technology. The first-order network is generally used to quickly generate candidate boxes for use in the later stage of the network. The general first-order network design method is to have a smaller resolution because of the speed of the top big funnel. It must be fast. For example, the input size of the face design is 12*12, but this design method will have problems in the actual use of engineering. The candidate frames generated by the first-order network will be very large, because the resolution is too small and the features are lost. , the ability to express is reduced, and sometimes there are as many as thousands of candidate boxes

From the perspective of the recall rate, the effect is good, but the effect of the error rate will be very poor, which will greatly affect the efficiency of the later stage. This is very fatal for embedded product applications, because there will be a lot of calculations, such as more Thousands of graphs are sent to the next-level network, and the next-level network is a refined network, the amount of calculation will increase a lot, greatly reducing the benefits of using cascade

Images