CN110674674A - A rotating target detection method based on YOLO V3 - Google Patents

Abstract

The invention discloses a rotating target detection method based on YOLO V3. The invention first extracts information from images of remote sensing ship training set images with arbitrary directions, uses the improved YOLO V3 algorithm to train the remote sensing ship training set, and finally trains the remote sensing ship training set. The detection of remote sensing ships with arbitrary angles can be done in the test set. The invention redesigns the generation of the anchor box, the calculation of the IOU, and the calculation method of the loss function of the original YOLO V3 algorithm, and increases the angle information of the detection target in the training set image, so that the target with any angle can be completed in the test set. detection.

Description

A rotating target detection method based on YOLO V3

technical field

The invention belongs to the field of deep learning, and relates to a rotating target detection method based on YOLO V3.

Background technique

At present, target detection has been widely used in military and civilian fields. A deep convolutional neural network can use the target dataset to autonomously learn the target to be detected and improve its own model. YOLO V3 is a single-step target detection algorithm. This algorithm does not need to use a regional candidate network (RPN) to extract candidate target information, but directly generates target location and category information through the network. It is an end-to-end target detection. algorithm. Therefore, the single-step object detection algorithm has a faster detection speed.

The YOLO V3 model uses the method of directly regressing the target coordinates and classification targets by dividing the grid for target detection. It mainly uses the horizontal rectangular bounding box to define the target position, and locates the target through the regression of the bounding box parameters. This method is accurate enough for targets such as people and vehicles, but it is not suitable for targets with arbitrary directions such as text and ships.

SUMMARY OF THE INVENTION

Aiming at the deficiencies of the prior art, the present invention provides a rotating target detection method based on YOLO V3. The method extracts information from images of remote sensing ship training sets with arbitrary directions, and uses the improved YOLO V3 algorithm to detect remote sensing ships. The training set is used for training, and finally the detection of remote sensing ships with arbitrary angles can be completed in the test set.

Step (1), extracting the position information and angle information of the detection target in the training set image;

The training set image is processed, and the detection target is represented as a 5-dimensional vector {x, y, w, h, θ}; where: x represents the coordinates of the detection target in the x-axis direction, and y represents the coordinates of the y-axis direction of the detection target. , w represents the length of the detection target, h represents the width of the detection target, and θ represents the inclination angle of the detection target;

Step (2), improve the YOLO V3 algorithm framework

2.1 Setting the anchor box

Use a three-dimensional variable anchor point with a rotation angle to set N different angles to control the extraction of the suspected area of the detection target; set the length and width of multiple sets of anchor boxes, and set three sets of anchors according to the known training set image detection target Point box size; different scales are generated according to the stride of the convolution kernel in the convolutional neural network and the size of the input image; use 3 anchors at each scale to generate 3 bounding boxes, resulting in 3*3*N An anchor box is used to fit the location area of the detection target;

2.2 Setting of loss function;

The delineation criteria for the extraction of positive and negative sample regions are: the IOU between the anchor box and the GT box is greater than 0.7, and the angle between the anchor box and the GT box is less than π/12, which is a positive sample; the IOU between the anchor box and the GT box is less than 0.3 or The IOU with the GT box is greater than 0.7, but the angle with the GT box is greater than π/12 is a negative sample; the samples that are not classified as positive samples and negative samples will not be used in the training process; the GT box is the location of the detection target area; IOU represents the overlapping area of the anchor box and the GT box fitted;

Define the loss function:

Loss=L _cls (p,l)+L _reg (v ^* ,v)+L _conf

Among them: L _cls (p, l) represents the loss function for classification, l represents the category label, l=1 means the target is selected, l=0 means the background is selected, if the background is selected, there is no regression loss function , when l=1, the loss of the real category and the predicted category is calculated by means of cross entropy, p is the probability of linear classification; L _reg (v ^* ,v) represents the loss function for regression; v ^* is step (1) The information of the detection target in (x, y, w, h, θ), v is the information of the predicted detection target corresponding to the anchor box in step (2) (x', y', w', h', θ') , the loss between the predicted value v and the real value v ^* is calculated by means of the mean square error; L _conf represents the loss function of the confidence degree. When the IOU is lower than 0.3, it is considered that there is no target in the picture. When it is higher than 0.3, the predicted value needs to be The target is as close to the real value as possible, and the part without the target needs to be as close to the real value of the background as possible, and the loss of confidence is calculated by means of cross entropy;

Step (3), use the improved YOLO V3 algorithm to train the training set images, iterate until the loss function no longer decreases, and obtain the weight file;

Step (4), use the weight file obtained in step (3) to test the test set images.

The beneficial effects of the present invention are as follows:

The invention redesigns the generation of the anchor box, the calculation of the IOU, and the calculation method of the loss function of the original YOLO V3 algorithm, and increases the angle information of the detection target in the training set image, so that the target with any angle can be completed in the test set. detection.

Description of drawings

FIG. 1 is a flow chart of the present invention.

Detailed ways

The present invention is further analyzed below in conjunction with FIG. 1 .

In this experiment, a group of collected ship target images are divided into training set and test set. The specific steps in the rotating target detection task based on YOLO V3 are as follows:

Step (1), extracting the position information and angle information of the detection target in the training set image.

Compared with the original YOLO V3 algorithm, the rotating target detection method needs to add an angle information. The training set images are processed to represent the detection target as a 5-dimensional vector {x, y, w, h, θ}. Where x represents the x-axis coordinate of the detection target, y represents the y-axis coordinate of the detection target, w represents the length of the detection target, w represents the width of the detection target, and θ represents the inclination angle of the detection target.

Step (2), improve the YOLO V3 algorithm framework.

2.1 Anchor box (anchor box) settings.

Use a three-dimensional variable anchor point with a rotation angle to set multiple different angles to control the extraction of the suspected area of the detection target. You can use 6 different angles: -π/6, 0, π/6, π/3, π/2, 2π/3. To set the length and width of multiple sets of anchor boxes, three sets of anchor box sizes can be set according to the known training set image detection targets: [(10, 13), (16, 30), (33, 23)], [(30 , 61), (62, 45), (59, 119)], [(116, 90), (156, 198), (373, 326)]. According to the stride of the convolution kernel in the convolutional neural network and the size of the input image, different scales will be generated. The stride of the original YOLO V3 feature map is 32, 16, and 8 respectively, and the size of the input image is 416*416. The resulting scale is 13*13, 26*26, 52*52. Use 3 anchors at each scale to generate 3 bounding boxes, so that 3*3*6=54 anchor boxes can be fitted to the location area of the detection target

2.2 Setting of loss function.

The delineation criteria for the extraction of positive and negative sample regions are: the IOU (the size of the overlap area fitted by the anchor box and the GT box) of the anchor box and the GT box (the location area of the detection target) is greater than 0.7, and the angle with the GT box is less than 0.7. π/12 is a positive sample; the IOU of the anchor box and the GT box is less than 0.3 or the IOU of the GT box is greater than 0.7, but the angle with the GT box is greater than π/12 is a negative sample; not classified as positive samples and negative samples Samples are not used during training.

Define the loss function:

Loss=L _cls (p,l)+L _reg (v ^* ,v)+L _conf

Among them: L _cls (p, l) represents the loss function for classification, l represents the category label (l=1 means the target is selected, l=0 means the background is selected, if the background is selected, there is no regression loss function ), when l=1, the loss between the real category and the predicted category is calculated by means of cross entropy (p is the probability of linear classification); L _reg (v ^* ,v) represents the loss function for regression. v ^* is the information of the detection target in step (1) (x, y, w, h, θ), v is the information of the predicted detection target corresponding to the anchorbox in step (2) (x', y', w', h) ', θ'), the loss between the predicted value v and the true value v ^* is calculated by means of the mean square error; L _conf represents the loss function of the confidence. When the IOU is lower than 0.3, it is considered that there is no target in the picture. When it is higher than 0.3, the predicted target needs to be as close to the real value as possible, and the part without the target needs to be as close to the real value of the background as possible. Cross entropy is used to calculate the confidence. degree loss.

Step (3): Use the improved YOLO V3 algorithm to train the training set images, iterate until the loss function no longer decreases, and obtain the weight file.

Step (4), use the weight file obtained in step (3) to test the test set images.

The above-mentioned embodiments are not intended to limit the present invention, and the present invention is not limited to the above-mentioned embodiments. As long as the requirements of the present invention are met, they all belong to the protection scope of the present invention.

Claims (1)

1. a rotating target detection method based on YOLO V3, is characterized in that, the method specifically comprises the following steps: Step (1), extracting the position information and angle information of the detection target in the training set image; The training set image is processed, and the detection target is represented as a 5-dimensional vector {x, y, w, h, θ}; where: x represents the coordinates of the detection target in the x-axis direction, and y represents the coordinates of the y-axis direction of the detection target. , w represents the length of the detection target, h represents the width of the detection target, and θ represents the inclination angle of the detection target; Step (2), improve the YOLO V3 algorithm framework 2.1 Setting of anchor box Use a three-dimensional variable anchor point with a rotation angle to set N different angles to control the extraction of the suspected area of the detection target; set the length and width of multiple sets of anchor boxes, and set three sets of anchors according to the known training set image detection target Point box size; different scales are generated according to the stride of the convolution kernel in the convolutional neural network and the size of the input image; use 3 anchors at each scale to generate 3 bounding boxes, resulting in 3*3*N An anchor box is used to fit the location area of the detection target; 2.2 Setting of loss function; The delineation criteria for the extraction of positive and negative sample regions are: the IOU of the anchor box and the GT box is greater than 0.7, and the angle between the anchor box and the GT box is less than π/12, which is a positive sample; the IOU of the anchor box and the GT box is less than 0.3 or The IOU with the GT box is greater than 0.7, but the angle with the GT box is greater than π/12, which is a negative sample; the samples that are not classified as positive samples and negative samples will not be used in the training process; the GT box is the location of the detection target area; IOU represents the overlapping area of the anchor box and the GT box fitted; Define the loss function: Loss=L _cls (p,l)+L _reg (v ^* ,v)+L _conf Among them: L _cls (p, l) represents the loss function for classification, l represents the category label, l=1 means the target is selected, l=0 means the background is selected, if the background is selected, there is no regression loss function , when l=1, the loss of the real category and the predicted category is calculated by means of cross entropy, p is the probability of linear classification; L _reg (v ^* ,v) represents the loss function for regression; v ^* is step (1) The information of the detection target in (x, y, w, h, θ), v is the information of the predicted detection target corresponding to the anchor box in step (2) (x', y', w', h', θ') , the loss between the predicted value v and the real value v ^* is calculated by means of the mean square error; L _conf represents the loss function of the confidence degree. When the IOU is lower than 0.3, it is considered that there is no target in the picture. When it is higher than 0.3, the predicted value needs to be The target is as close to the real value as possible, and the part without the target needs to be as close to the real value of the background as possible, and the loss of confidence is calculated by means of cross entropy; Step (3), use the improved YOLO V3 algorithm to train the training set images, iterate until the loss function no longer decreases, and obtain the weight file; Step (4), use the weight file obtained in step (3) to test the test set images.

Images