An Aircraft Target Recognition Method Based on Circular Filtering and Convolutional Neural Network

Abstract

The present invention relates to the technical field of target recognition, and particularly provides an aircraft target recognition method based on circular filtering and convolutional neural network. By using the shape and structure characteristics of the aircraft itself, circular filtering features are extracted for feature enhancement, combined with the features of the deep neural network, and finally, aircraft target recognition based on circular filtering and convolutional neural network is learned. The present invention has rotational invariance and scale invariance, and has more accurate positioning compared with the recognition using only convolutional neural network, can accurately recognize aircraft targets, and has a high recognition accuracy and recall rate.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition, and particularly provides an aircraft target recognition method based on circular filtering and convolutional neural network. Background Art

[0002] Aircraft targets are important transportation carriers and high-value targets both in the civilian and military fields. In recent years, deep learning has achieved excellent performance in various fields of target recognition. However, there are still many problems in the recognition of aircraft targets with complex backgrounds and different models, sizes, and postures. Therefore, the invention uses the shape and structure characteristics of the aircraft itself for circular filtering, combines the characteristics of the deep neural network, makes the positioning of the aircraft target more accurate, and improves the recognition accuracy and recall rate of the aircraft target. Summary of the Invention

[0003] The purpose of the present invention is to provide an aircraft target recognition method based on circular filtering and convolutional neural network to solve the problem of recognizing aircraft targets with complex backgrounds and different models, sizes, and postures.

[0004] The technical solution adopted by the present invention to achieve the above purpose is as follows:

[0005] An aircraft target recognition method based on circular filtering and convolutional neural network includes the following steps:

[0006] Step 1: Input the original image with an aircraft.

[0007] Step 2: Extract circular filtering features from the input original image.

[0008] Step 3: Enhance the features of the extracted circular filtering features on the channels of the original image.

[0009] Step 4: Input the enhanced image into the deep target detection network, and use the backpropagation algorithm to update the network parameters to train the deep target detection network.

[0010] Step 5: Input the image to be detected into the trained deep target detection network and output the recognition result.

[0011] The specific content of step 2 is as follows:

[0012] Select the pixel values on the circumferences of multiple concentric circles with the center of the aircraft as the center and a radius greater than the fuselage width and less than the wingspan length. After discrete Fourier transform, the amplitude response is obtained, and the result after averaging the amplitude responses of multiple circumferences with different radii is used as the circular filtering feature map.

[0013] The pixel value column on the circumference is transformed by discrete Fourier transform as follows to obtain the amplitude response:

[0014]

[0015] where f k represents the gray value of the pixels in the counterclockwise direction on the circumference centered at (i, j), c is the period, and N is the total number of pixels on the circumference.

[0016] Specifically, in step 3: the circular filtering feature map obtained in step 2 is superimposed on the original image as the first channel of the feature-enhanced image, the original image is used as the second channel of the feature-enhanced image, and the circular filtering feature map is used as the third channel of the feature-enhanced image.

[0017] The deep object detection network includes a backbone network, a neck network, and a convolutional prediction module connected in sequence, where: the backbone network first performs Mosaic data augmentation on the input image, and then performs interlaced sampling on it to split and re-stack the high-resolution input image into multiple low-resolution image channels through interlaced sampling; then the processed feature image passes through the backbone network basic module and the CSP structure in sequence, and the output result is used as the input of the neck network;

[0018] The neck network adopts the FPN+PAN structure. The FPN layer conveys semantic features from top to bottom, and the PAN layer conveys localization features from bottom to top. The output of the neck network is predicted through the convolutional prediction module;

[0019] The convolutional prediction module identifies aircraft targets through convolutional layers.

[0020] The backbone network basic module includes a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence.

[0021] Updating the network parameters using the backpropagation algorithm to train the deep object detection network is specifically as follows: the predicted bounding boxes output by the deep object detection network are compared with the ground truth boxes, the difference between the two is calculated through the loss function, and then updated backward. The network parameters are iteratively trained. When the network loss is less than the set value, the iteration stops and the network training is completed.

[0022] The network loss function is as follows:

[0023]

[0024] S is the size of the feature map cell obtained finally. The feature map is divided into S×S cells, and B is the number of predicted bounding boxes for each cell; indicates whether the center coordinates of the target fall into cell j. If so, the value is 1; otherwise, the value is 0; p j is the probability that this predicted bounding box belongs to class C, classes represents the set of classes in the dataset, and BCE_loss is the binary cross-entropy loss function; Indicates whether the k-th prediction box can represent the prediction result of cell j when the target center falls within cell j. If this prediction box has the highest Intersection over Union (IOU) value with the ground truth box among all B prediction boxes of cell j, it can represent, and the value is 1 if it can represent, otherwise 0. Indicates whether the k-th prediction box can represent the prediction result of cell j when the target center does not fall within cell j. If this prediction box has the highest IOU value with the ground truth box among all B prediction boxes of cell j, it can represent, and the value is 1 if it can represent, otherwise 0; conf j Is the confidence of this prediction box. The letter with a hat symbol above is the predicted value, and the one without the hat symbol is the true value.

[0025] The position error is as follows:

[0026]

[0027] Where ρ(b p , b gt ) is the Euclidean distance between the center point coordinates of the prediction box and the ground truth box; d is the diagonal distance of the smallest bounding rectangle enclosing them.

[0028] Step 5 is specifically as follows:

[0029] Input the image to be detected into the deep object detection network with updated network parameters, perform non-maximum suppression on all detected prediction boxes, and finally output the aircraft target recognition result.

[0030] The present invention has the following beneficial effects and advantages:

[0031] The present invention has rotational invariance and scale invariance. Compared with using only a convolutional neural network for recognition, the positioning is more accurate, it can accurately recognize aircraft targets, and has a high recognition accuracy rate and recall rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Is a schematic flow chart of an aircraft target recognition method using circular filtering and a convolutional neural network provided by the present invention;

[0033] Figure 2a Is a schematic diagram of the result of circular filtering feature extraction in an aircraft target recognition method using circular filtering and a convolutional neural network provided by the present invention;

[0034] Figure 2b Is a schematic diagram of the result of circular filtering feature extraction in an aircraft target recognition method using circular filtering and a convolutional neural network provided by the present invention;

[0035] Figure 3Schematic diagram of the recognition result of the test set of an aircraft target recognition method based on circular filtering and convolutional neural network provided by the present invention. Detailed implementation manners

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0037] The present invention discloses an aircraft target recognition method based on circular filtering and convolutional neural network. The method process includes five parts: original image input, circular filtering feature extraction, deep target detection network construction, network update and optimization, and recognition result output.

[0038] An aircraft target recognition method based on circular filtering and convolutional neural network, comprising:

[0039] Step 1: Input the original image;

[0040] Step 2: Extract circular filtering features from the input original image;

[0041] Step 3: Enhance the features of the extracted circular filtering features on the channels of the original image;

[0042] Step 4: Input the enhanced image into the deep target detection network;

[0043] Step 5: Update the network parameters using the backpropagation algorithm;

[0044] Step 6: Output the recognition result.

[0045] An aircraft has relatively fixed geometric shape features. An aircraft is mainly composed of a fuselage and two wings on both sides. If a circle is selected with the center of the aircraft as the center and the radius is greater than the width of the fuselage and less than the wingspan length, then no matter moving clockwise or counterclockwise on this circle, it will pass through 8 parts that are alternately transformed between the aircraft and the background parts divided by the fuselage, tail, and two wings on both sides. Whether in infrared images or visible light images, the aircraft target often has a certain contrast with the background. Therefore, the gray level of the circular pixels will show a changing trend of 4 peaks and 4 valleys, similar to a sine / cosine function with 4 periods, and has rotational invariance.

[0046] As Figures 2a to 2b shown, where Figure 2a is the aircraft image after gray level transformation in the satellite image, Figure 2b is the circular filtering feature extracted from Figure 2a . Step 2 includes obtaining the amplitude response after discrete Fourier transform of the circular pixel values selected with the center of the aircraft as the center and the radius greater than the width of the fuselage and less than the wingspan length. To adapt to aircraft targets of different scales, the result of averaging multi-scale circular filtering is taken, and finally a circular filtering feature map is obtained.

[0047] In step 3, it includes superimposing the circular filtering feature map obtained in step 2 onto the channels of the input original image for feature enhancement.

[0048] In step 4, it includes performing deep neural network feature extraction on the input image after feature enhancement.

[0049] In step 5, it includes performing backpropagation update according to the training data labels and iteratively updating the network parameters. When the network loss is less than the set value or the loss curve has tended to converge, the training is completed.

[0050] In step 6, it includes inputting the image to be detected into the deep object detection network after updating the network parameters, performing non-maximum suppression on all detected prediction boxes, and finally outputting the aircraft target recognition result.

[0051] As Figure 1 shown, the present invention provides an aircraft target recognition method based on circular filtering and convolutional neural network, including the following steps:

[0052] Step 1: An aircraft target data set containing n pictures x i is the i-th image data input into the network, and the label label i (ground truth box) contains the center position coordinates, width and height of the target, and class attribution of all aircraft targets in this image.

[0053] Step 2: For the training input image x i Three different scales of circles are selected with the center of the aircraft as the center, greater than the fuselage width and less than the wingspan length. For the pixel values of each scale of circle column, the amplitude response is obtained after performing the discrete Fourier transform according to the following formula:

[0054]

[0055] where f k represents the gray value of the pixels in the counterclockwise direction on the circle centered at (i, j), and c takes 8. The result f(i, j) after averaging the multi-scale circular filtering is normalized to adapt to aircraft targets of different scales, and finally the feature map c after circular filtering of the original image is obtained i .

[0056] Step 3: Perform feature enhancement on the channels of the original image for the extracted circular filtering features, and superimpose c i onto the first and third channels of x i and perform a normalization operation to obtain x i '.

[0057] Step 4: For the input image x after feature enhancement i'Then perform Mosaic data augmentation, and perform multi-scale training after splicing the images in a random scaling, cropping, and arranging manner. For x i 'Interlaced sampling is used to split and then re-stack into multiple low-resolution image channels. For example, an image of 608×608×3 is converted into a feature map of 304×304×12, and then finally becomes a feature map of 304×304×32 through convolution, reducing the information loss caused by downsampling. Next, the basic module of the backbone network includes a convolutional layer, batch normalization (Batch Normalization), and a Leaky Relu activation function. Then, a CSP structure is connected. By using the splitting and fusion strategies, the gradient path is increased, and the gradient flow is truncated to prevent different layers from learning duplicate gradient information; features are integrated across stages, enabling the features to be reused, balancing the computational amount of each layer, and reducing memory occupancy.

[0058] The neck network adopts the FPN+PAN structure. The FPN layer conveys strong semantic features from top to bottom, while the PAN conveys strong localization features from bottom to top, further improving the ability of feature extraction. At the same time, the CSP2 structure designed by CSPNet is used to strengthen the ability of network feature fusion.

[0059] Step 5: The network adaptively calculates the best prior values in different training sets, outputs prediction boxes on the best boxes, and then compares them with the ground truth of the real boxes. The network loss function consists of three parts: position error, confidence error, and classification error. Calculate the error between the real label and the prediction result. The specific network loss function is as follows:

[0060]

[0061] S is the size of the cell of the finally obtained feature map. The feature map is divided into S×S cells. B is the number of prediction boxes for each cell; λ noobj is the weight parameter, indicates whether the center coordinates of the target fall within cell j. If so, the value is 1; otherwise, the value is 0; p j is the probability that this prediction box belongs to class C. classes represents the set of classes in the dataset. BCE_loss is the binary cross-entropy loss function; indicates whether the k-th prediction box can represent the prediction result of cell j when the target center falls within cell j. If this prediction box is the one with the highest IOU value with the real box among all B prediction boxes of cell j, then it can represent, and the value is 1 if it can represent; otherwise, the value is 0; indicates whether the k-th prediction box can represent the prediction result of cell j when the target center does not fall within cell j. If this prediction box is the one with the highest IOU value with the real box among all B prediction boxes of cell j, then it can represent, and the value is 1 if it can represent; otherwise, the value is 0; confj is the confidence of the prediction box. The letter with a roof symbol above is the predicted value, and the letter without the roof symbol is the true value.

[0062]

[0063] Among them, ρ(b p , b gt ) is the Euclidean distance between the center point coordinates of the prediction box and the true box; d is the diagonal distance of the smallest bounding rectangle that encloses them;

[0064] At the same time, it measures the consistency of the width-to-height ratio of the prediction box and the true box:

[0065]

[0066] w p represents the width of the prediction box, h p represents the height of the prediction box, w gt represents the width of the prediction box, h gt represents the height of the prediction box.

[0067] By it avoids generating a large bounding rectangle when the distance between the prediction box and the true box is far, resulting in a large loss value that is difficult to optimize. Even in the case where one box contains the other box, effective measurement can be performed, and the convergence speed is accelerated. is used to ensure that when the center points of the prediction box and the true box coincide, effective measurement is performed through the width-to-height ratio of the box.

[0068] Through backpropagation using the gradient descent method, the network parameters are iteratively updated, making the network loss gradually decrease. When the network loss is less than the set value or the loss curve has converged, the training is completed.

[0069] Step 6: Scale the image i to be detected, adaptively add the least amount of black edges to accelerate the inference running speed, and then input it into the deep object detection network with updated network parameters. For all the detected prediction boxes, non-maximum suppression is performed, and finally the aircraft target recognition result is output.

[0070] The final detection result of the aircraft target recognition method based on circular filtering and convolutional neural network proposed by the invention is as Figure 3 shown. The square box is the detected aircraft target. The overall recognition has a precision of 92.9% and a recall rate of 91.9% under the condition of IOU = 0.65.

[0071] Among them, the definition of precision:

[0072]

[0073] The definition of recall rate:

[0074]

[0075] Table 1 Experimental results of the present invention

[0076]

[0077] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those of ordinary skill in the art.

Claims

1. An aircraft target recognition method based on circular filtering and convolutional neural network, characterized in that It includes the following steps: Step 1: Input the original image with an aircraft; Step 2: Extract the circular filtering features from the input original image; Step 3: Enhance the features of the extracted circular filtering features on the channels of the original image; Step 4: Input the enhanced image into the deep object detection network, and use the backpropagation algorithm to update the network parameters to train the deep object detection network; Step 5: Input the image to be detected into the trained deep object detection network and output the recognition result; The specific content of Step 2 is as follows: Select the pixel values on the circumferences of multiple concentric circles with the center of the aircraft as the center and a radius greater than the fuselage width and less than the wingspan length. After discrete Fourier transform, obtain the amplitude response, and take the average result of the amplitude responses of multiple circumferences with different radii as the circular filtering feature map; The deep object detection network includes a backbone network, a neck network, and a convolutional prediction module connected in sequence, where: the backbone network first performs Mosaic data augmentation on the input image, and then performs interlaced sampling on it to split the high-resolution input image into multiple low-resolution image channels after interlaced sampling and re-stack them; then the processed feature image passes through the backbone network basic module and the CSP structure in sequence, and the output result is used as the input of the neck network; The neck network adopts the FPN+PAN structure. The FPN layer conveys semantic features from top to bottom, and the PAN layer conveys localization features from bottom to top. The output of the neck network is predicted through the convolutional prediction module; The convolutional prediction module identifies the aircraft target through the convolutional layer.

2. The aircraft target recognition method based on circular filtering and convolutional neural network according to claim 1, characterized in that, The pixel value column on the circumference obtains the amplitude response after discrete Fourier transform through the following formula: where f k represents the gray value of the pixels in the counterclockwise direction on the circumference centered at (i, j), c is the period, and N is the total number of pixels on the circumference.

3. The aircraft target recognition method based on circular filtering and convolutional neural network according to claim 1, wherein, The specific content of Step 3 is as follows: Superimpose the circular filtering feature map obtained in Step 2 on the original image as the first channel of the feature-enhanced image, use the original image as the second channel of the feature-enhanced image, and use the circular filtering feature map as the third channel of the feature-enhanced image.

4. The aircraft target recognition method based on circular filtering and convolutional neural network according to claim 1, wherein, The backbone network basic module includes a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence.

5. A method for aircraft target recognition based on circular filtering and convolutional neural network according to claim 1, characterized in that The use of the backpropagation algorithm to update the network parameters to train the deep object detection network is specifically as follows: Compare the predicted bounding boxes output by the deep object detection network with the ground truth boxes, calculate the difference between the two through the loss function, and then update backward and iteratively train the network parameters. When the network loss is less than the set value, stop the iteration and the network training is completed.

6. The aircraft target recognition method based on circular filtering and convolutional neural network according to claim 5, characterized in that, The network loss function is as follows: S is the size of the feature map cell obtained finally. The feature map is divided into S×S cells, and B is the number of prediction boxes for each cell; Indicates whether the center coordinates of the target fall within cell j. If so, the value is 1; otherwise, the value is 0; p j (C) is the true probability value that the prediction box belongs to class C. is the predicted probability value that the prediction box belongs to class C. classes represents the set of classes in the dataset, and BCE_loss is the binary cross-entropy loss function; Indicates whether the k-th prediction box can represent the prediction result of cell j when the target center falls within cell j. If this prediction box has the highest IOU value with the true box among all B prediction boxes in cell j, it can represent, and the value is 1 if it can represent; otherwise, the value is 0; Indicates whether the k-th prediction box can represent the prediction result of cell j when the target center does not fall within cell j. If this prediction box has the highest IOU value with the true box among all B prediction boxes in cell j, it can represent, and the value is 1 if it can represent; otherwise, the value is 0; conf j is the true confidence value of this prediction box. is the predicted confidence value of this prediction box, λ noobj is the weight parameter.

7. The aircraft target recognition method based on circular filtering and convolutional neural network according to claim 6, characterized in that, The position error is: Among them, ρ(b p ,b gt ) is the Euclidean distance between the center point coordinates of the predicted box and the ground truth box; d is the diagonal distance of the smallest bounding rectangle enclosing them, and ν represents the aspect ratio consistency between the predicted box and the ground truth box.

8. A method for aircraft target recognition based on circular filtering and convolutional neural network according to claim 1, characterized in that, The specific content of Step 5 is as follows: Input the image to be detected into the deep object detection network with updated network parameters, perform non-maximum suppression on all detected predicted bounding boxes, and finally output the aircraft target recognition result.

Images