Target detection model training method and device based on deep learning and storage medium
A technology for target detection and model training, applied in the field of deep learning target detection, can solve problems such as low confidence and wrong category labeling, achieve good performance, and improve accuracy and recall.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Example Embodiment
[0027] Example one:
[0028] First, refer to figure 1 Describe the deep learning-based target detection model training method of the present invention. Such as figure 1 As shown, the training method of the target detection model based on deep learning includes the following steps:
[0029] S101: Perform a test on training images containing target annotations to obtain targets contained in each training image. Specifically, when testing training images containing target annotations, you can use external public data sets (such as COCO data sets, etc.) to test the training images; or directly input the training images to the trained target detection In the model, the training image is tested through the target detection model. At this time, the confidence threshold will be set lower to reduce the possibility of missing certain targets. Specifically, the threshold can be set according to the recall rate of the algorithm, such as the confidence level corresponding to the recall rate o...
Example Embodiment
[0036] Embodiment two:
[0037] When training the target detection model, it is necessary to calculate the true value (background or a certain category) of each position in the training image. In the anchor-based algorithm, it is usually to calculate the intersection of the bounding box (anchor) of the target label and the prediction box (Intersection over Union, iou), if iou is greater than a preset threshold, it is set as a positive sample, otherwise it is Negative sample. Due to the influence of the target's translation and size, the iou of some targets and anchors may be smaller than the preset threshold. If the positive and negative samples are divided directly according to the iou, the positive samples will be missed. In the anchor-free algorithm, when it is judged that the size of the target is within the scale range of the network of this layer, and the characteristic point in the bounding box of the target label is found, the characteristic point is set to a certain cla...
Example Embodiment
[0042] Example three:
[0043] In this embodiment, a further design is made on the basis of the second embodiment, which takes into account the similarity of the categories when calculating the classification loss of the positive samples. For example, when the target is a pedestrian, the probability that the cyclist will be judged as the background will be lower.
[0044] Specifically, first, set the similarity matrix of each category, the value of each element of the main diagonal in the similarity matrix is 1, and the other elements are based on the category similarity in the interval [0,1], the categories are similar The greater the degree, the smaller the value of the element to reduce the mutual influence between the elements. Then, the classification loss of each category is calculated for each feature point, and finally the products are multiplied with the similarity matrix and added to obtain the classification loss of the positive sample. For example, the similarity ma...
PUM
Abstract
Description
Claims
Application Information
- R&D Engineer
- R&D Manager
- IP Professional
- Industry Leading Data Capabilities
- Powerful AI technology
- Patent DNA Extraction
Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.
© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap