Single-stage arbitrary quadrilateral regression frame large length-width ratio target remote sensing image detection algorithm
A technology of remote sensing images and detection algorithms, applied in the field of remote sensing images, can solve the problems of accurate positioning of targets with large aspect ratios and the inability to use horizontal frames, etc., and achieve the effect of rapid detection
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
specific Embodiment approach 1
[0062] Specific implementation mode one: the following combination figure 1 Describe this embodiment, the single-stage arbitrary quadrilateral regression frame large aspect ratio target remote sensing image detection algorithm described in this embodiment, the detection algorithm is based on the single-stage target detection framework, and can return any quadrilateral;
[0063] The specific process includes:
[0064] S1. Using the feature pyramid network structure, perform feature extraction on the three feature layers of the target remote sensing image, and fuse the extracted features;
[0065] S2. Using any quadrilateral frame to perform regression calculation on the target position of the target remote sensing image, obtain a candidate frame of any quadrilateral, and obtain classification results and confidence scores at the same time;
[0066] S3. Merge the candidate frames with high confidence scores on the three scales, restore them to the original size, calculate the i...
specific Embodiment approach 2
[0067] Specific implementation mode two: this implementation mode further explains specific implementation mode one, the feature extraction of the three feature layers of the target remote sensing image described in S1 is performed using the CSP-Darknet53 network for calculation;
[0068] Specifically include:
[0069] When performing feature extraction on the deep feature map, copy the feature map of the base layer;
[0070] When performing feature extraction on feature maps of different scales, the upper-layer information and lower-layer information are respectively combined with up-down sampling.
[0071] In this embodiment, traditional target detection algorithms are generally divided into single-stage and two-stage algorithms. The two-stage algorithm first extracts the region of interest, and then performs classification and regression calculations on each candidate region. Although this method improves the detection accuracy to a certain extent, it greatly increases th...
specific Embodiment approach 3
[0072] Embodiment 3: This embodiment further explains Embodiment 1 or 2. The regression calculation in S2 includes: regression center coordinates, regression width and height, and addition of four offsets to realize regression rotation.
[0073] In this embodiment, since the remote sensing image is different from the traditional image, the objects are all from a bird's-eye view, so the target with a large aspect ratio, such as a ship, cannot be accurately positioned with a horizontal frame. The use of any quadrilateral frame in S2 can realize the regression calculation of the inclined target position. In the regression part, in addition to returning to the conventional center coordinates and width and height, four additional offsets are added to achieve rotation regression. In addition, the present invention does not use angle as a regression parameter. On the one hand, it avoids the periodic problem caused by angle regression. On the other hand, it uses four parameters to des...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More