Priori box design method in target detection algorithm based on anchor box

Abstract

The invention discloses a priori box design method in a target detection algorithm based on an anchor box, and the method is a priori box design method based on a training set sample target marking box area clustering algorithm, and comprises the steps: carrying out the two rounds of clustering algorithm calculation on the whole, and obtaining the area hyper-parameter of a priori box through the first round of clustering calculation; and performing a second round of clustering according to the result to determine the length-width ratio hyper-parameter of the priori frame, thereby finishing the deployment of the priori frame. According to the method, the problem of specific data imbalance of a target detection algorithm in an actual application scene is fully considered, that is, the number of small target samples in a data set is one order of magnitude or more than one order of magnitude greater than the number of large target samples, and a clustering algorithm based on a mark box area is innovatively used; the defect that the number of samples is small is balanced by using the characteristic that the area value of a large target is large, so that priori frame hyper-parameters more suitable for a scene with large and small targets are obtained, the training difficulty of a neural network model can be reduced, the training process is accelerated, and a target detection model with a better effect is obtained.

Description

technical field [0001] The invention belongs to the anchor frame-based neural network target detection technology, in particular to a priori frame design method in the anchor frame-based target detection algorithm. Background technique [0002] At present, in the actual implementation of the target detection algorithm, the target detection algorithm based on the anchor frame has the advantage of low time consumption due to its single-stage paradigm, which can meet the real-time requirements of the project, so this type of algorithm is applied in the actual implementation widely. [0003] In the practical application of this kind of target detection algorithm based on the anchor box, it is necessary to set the hyperparameters of the network prior frame according to the actual application scene or the training set. Shift, and then get the position of the target after decoding to complete the target detection task. The final model effect of training depends largely on the set...

AI Technical Summary

Problems solved by technology

When the target detection algorithm is actually implemented, it is difficult to ensure sample balance in data set collection, and there will be more or less sample imbalance.

Images