Cross-domain person re-identification method based on normalization and feature enhancement
A pedestrian re-identification and feature enhancement technology, applied in the field of pedestrian re-identification, can solve the problems of wasting computing resources, redundancy, etc., and achieve the effects of enhancing pedestrian characteristics, reducing distance, and enhancing pedestrian discrimination features
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0054] This embodiment discloses a cross-domain pedestrian re-identification method based on normalization and feature enhancement. As a basic implementation of the present invention, it includes establishing a recognition network model, normalizing image features, restoring image features, and outputting image features.
[0055] Build a recognition network model, including building such as figure 2 shown with an example normalization unit IN (ie figure 2 IN in), the residual weight training unit CMS (ie figure 2 CMS in ) and attention unit CAB (ie figure 2 The normalized enhancement module NEM of CA in CA), and the ResNet50 model is used as the backbone network, and the normalized enhanced module NEM is inserted into the ResNet50 model to form a recognition network model.
[0056] Image feature normalization, which normalizes the style of features by calculating the mean and variance in each channel of image features, which can suppress style differences between differe...
Embodiment 2
[0077]This embodiment discloses a cross-domain pedestrian re-identification method based on normalization and feature enhancement, as a preferred implementation of the present invention, that is, in Example 1, the ResNet50 model includes a Res1 unit, a Res2 unit, and a Res3 unit that are sequentially connected by communication. Unit, Res4 unit, Res5 unit and Head unit, insert the normalized enhancement module NEM after each Res unit or part of the Res unit of the ResNet50 model, which can enhance the features in the relevant stages respectively, so that the overall effect is good. of. In the ResNet50 model, the features obtained by the Res1 unit operation are too shallow and basically do not contain semantic information such as style. Inserting the normalized enhancement module NEM after the Res1 unit will not enhance the feature and will further increase the complexity of the model. Therefore, in the process of designing the recognition network model, the normalization enhanc...
Embodiment 3
[0080] This example discloses a cross-domain pedestrian re-identification method based on normalization and feature enhancement, as a preferred implementation of the present invention, that is, in Example 2, in order to promote features with better clustering characteristics, it is also included in The NEM loss function is introduced into the normalized enhancement module NEM at the output of the Res5 unit, and the normalized enhancement module NEM (ie NEM5) is constrained, and the features extracted by NEM5 are expected to have better domain invariance and discriminative power, Therefore, the image feature output of the normalized enhancement module NEM also includes the following steps:
[0081] S33, respectively calculate the center loss C of the input feature x x and the output feature f center loss C f , to measure the intra-class dispersion of the input feature x and the output feature f in the feature space, the calculation formula is:
[0082]
[0083]
[0084]...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More