Position classification method and device of target detection network and electronic equipment

Abstract

The invention provides a position classification method and device of a target detection network, and electronic equipment. The method comprises the following steps: dividing parameters used for determining a target frame in the target detection network into classification vectors according to a preset length; and obtaining response values of the classification vectors, and determining a position parameter corresponding to the maximum response value in the classification vectors or a position parameter obtained after mathematical transformation of the position parameter as a target frame parameter of the target detection network. The objective of the invention is to overcome the defect of large error when a regression method is used to predict a coordinate frame under low-bit quantization in the prior art, realize that a coordinate regression part of a target detection network is realized by a classification method, and greatly improve the precision of the network after overall quantization on the premise of ensuring the effect of a full-precision network.

Description

technical field [0001] The invention relates to the technical field of target detection with neural network algorithms, in particular to a position classification method, device and electronic equipment of a target detection network. Background technique [0002] Object detection is an important research direction in the field of neural networks. It is widely used in various industries. The better the effect of the neural network, the more complex the model, which has caused great difficulties for mobile terminal deployment. Therefore, many model compression methods for neural networks have emerged, among which quantization is a more commonly used method. However, neural networks for different tasks have their own characteristics. The current quantification methods mostly design experiments in classification networks. Even if good results are achieved, they are only for classification networks. [0003] The target detection task is more complicated than the classification t...

AI Technical Summary

Problems solved by technology

[0004] At present, target detection commonly uses regression methods to predict coordinate frames, that is, as figure 1 As shown, the input features figure 1 The coordinate offset value 3 (tx, ty, tw, th) is calculated after passing through the neural network 2, where tx and ty are the offset values ​​in the horizontal and vertical directions of the coordinates of the center point of the target frame, respectively, and tw and th are respectively is the length and width offset values ​​of the target frame, and the real target frame parameters can be obtained after (tx, ty, tw, th) is subjected to nonlinear mathematical transformation and scaling and translation in the subsequent processing. However, the network In the case of low-bit quantization, the nonlinear transformation on the target box size will amplify the error

Images