Small target detection method based on improved YOLOv3

Abstract

The invention discloses a small target detection method based on improved YOLOv3, and the method comprises the steps: constructing a feature extraction network of the improved YOLOv3, and replacing the DarkNet-53 of the YOLOv3 with the feature extraction network which consists of residual modules containing a channel attention mechanism; constructing a feature fusion network of the improved YOLOv3, and adding a structure from bottom to top on the basis of a feature pyramid network of the YOLOv3; adopting a K-means + + clustering algorithm to replace the K-means algorithm to cluster the self-made data set to generate an anchor box; constructing an improved YOLOv3 detection network, and carrying out detection on feature maps with the sizes of 52 * 52 and 26 * 26; and constructing a loss function of the improved YOLOv3, and the CIoU loss function being used as the loss of the target box. Compared with an existing YOLOv3 algorithm, the method has the advantages that the detection precisionof the small target is improved by 9.96%, and the detection precision is obviously improved.

Description

technical field [0001] The invention mainly relates to the field of computer vision, in particular to a detection method of a small target. Background technique [0002] At present, the target detection technology based on deep learning has achieved great success and is widely used in real life. However, this technology is mainly aimed at relatively large targets. In the small target detection task (in the MS COCO dataset, 32×32 pixels or less are defined as small targets), on the one hand, due to the small size of the small target, it has information It is also less, and the information after the multi-layer convolutional neural network is less. On the other hand, the current small target data set is also small, which cannot provide support for the training of the deep model, resulting in poor detection effect of small targets. Omissions occur. Contents of the invention [0003] A kind of small target detection method based on improved YOLOv3 of the present invention, t...

AI Technical Summary

Problems solved by technology

However, this technology is mainly aimed at targets with relatively large sizes. In the small target detection task (in the MS COCO dataset, small targets below 32×32 pixels are defined), on the one hand, due to the small size of small targets, they have information It is also less, and the information after the multi-layer convolutional neural network is less. On the other hand, the current small target data set is also small, which cannot provide support for the training of the deep model, resulting in poor detection effect of small targets. Occurrence of missed detection

Images