foreword
I published the target improvement algorithm column before, but for what scenario to apply, what improvement method is needed to be effective for your own application scenario, and how many improvements can be published at what level of articles, in order to solve everyone's confusion, this series of articles aims to Interpret the latest target detection algorithm papers for everyone to help you answer your doubts. In the series of articles interpreted, I have reproduced the innovative code, and friends in need can pay attention to private messages.
1. Summary
Target detection is an important technology in the field of remote sensing image processing. There are many types of targets in remote sensing images and there are target objects that are difficult to detect. A method of applying the YOLOv5 algorithm to remote sensing image target detection is proposed. Firstly, YOLOv5x is selected to construct the network model, and then the sample set is preprocessed through Mosaic data enhancement and the adaptive anchor frame screening method is used to determine the size of the anchor frame, and then the slice convolution operation is obtained. The original feature map is sent to the backbone network for feature fusion to obtain the optimal weight. Finally, GIOU Loss is used as the loss function of the bounding box and the screening of the non-maximum suppression target box to perform target detection on remote sensing images. Detection experiments are conducted on a public dataset of 10 types of geospatial objects (NWPU-VHR10) to evaluate the object detection performance of the proposed model. Comparative experiments show that the mAP of the model in this paper reaches 0.9239. Compared with the best results in the model using the same data set, the mAP is increased by 1.78%. This method can improve the accuracy of target detection in remote sensing images.
2. Network model and core innovation points
1. Mosaic data enhancement
2. Adaptive methods
3. GIOU Loss function
4. Non-Maximum Suppression (NMS)
3. Application data set
The NWPU-VHR10 dataset is a remote sensing image collected from Google Earth and Vaihingen datasets, and then the target objects are marked by experts. The dataset consists of 800 images, including a set of 650 positive images that contain at least one object to be detected, a set of 150 negative images that do not contain any object of a given object class, and 80 remote sensing images for each class.
4. Experimental results (partial display)
In order to evaluate the detection performance of the model, this paper randomly selects remote sensing data sets with different training and testing ratios, uses different ratios of image data set samples to train the model, obtains 9 optimal weight values, and conducts 10 groups of different ratios of different test sets. experiment. By calculating the mAP results of 10 groups of different tests, the arithmetic mean mAP value is finally obtained.
- The figure below shows the recall curves for different ratios during object detection.
2. The table below shows the arithmetic mean mAP values for experiments that adjusted the dataset according to these ratios.
5. Experimental conclusion
The experimental results show that the application of YOLOv5 algorithm can quickly detect remote sensing image targets, and the average precision and average precision have been significantly improved, and 9 sets of data sets with different ratios are used, and the verified model has good robustness.
6. Introduction to journals
Note: The original text of the paper is from Xing Yuchi, Li Dajun, Ye Famao. Target detection in remote sensing images based on YOLOv5. Jiangxi Science. Volume 39, Issue 4, August 2021.
For the series of articles interpreted, I have reproduced the innovative code. Friends in need can pay attention to the following official account and private message me.