An angle-based target detection training method, target detection method and device

Abstract

Aspects of the present invention relate to the field of image processing, and provide an angle-based target detection training method, target detection method and device. An angle-based target detection training method, which includes: obtaining an image step, obtaining a training image; obtaining a prediction frame step, obtaining a prediction frame through a neural network according to the obtained training image; obtaining a marked point step, obtaining the first vertex of the label frame , the second vertex of the prediction frame corresponding to the position of the first vertex, and the third point in the label frame; the step of obtaining the loss is to obtain the loss based on the target loss, wherein the target loss includes the angle loss, and the angle loss is based on the first vertex and the second The angle between the line connecting the three points and the line connecting the second vertex and the third point is obtained; the optimization step is to train the neural network based on the target loss. The target detection process is constrained by introducing an angle loss function, which is used to enhance the constraint of the overlapping degree between the labeled frame and the predicted frame, thereby improving the accuracy of target detection.

Description

technical field [0001] The present invention generally relates to the field of image processing, in particular to an angle-based target detection training method, target detection method and device. Background technique [0002] Object detection is an important research direction in the field of computer vision. How to improve the performance of object detectors has always been a topic that many researchers want to break through. At present, the target loss and classification loss commonly used in target detection are the smoothed Manhattan distance loss function (SmoothL1) and the normalized exponential function (Softmax), where SmoothL1 measures the distance between the prediction frame and the label frame. However, in practical applications, the SmoothL1 loss function cannot fully express the strong correlation between the prediction frame and the annotation frame. Usually, the distance between the prediction frame and the annotation frame detected by the distance loss fu...

AI Technical Summary

Problems solved by technology

However, if IoU is used as the loss function, the difference between the prediction frame and the annotation frame in different alignment directions cannot be distinguished. At the same time, when the prediction frame and the annotation frame do not overlap, the gradient will be 0, which will cause the network to be unable to optimize. Measure the distance between the predicted box and the labeled box

Images