Polarization imaging camouflage target identification method based on deep learning

Abstract

The invention provides a polarization imaging camouflage target identification method based on deep learning. Comprising the following steps: S1, obtaining 0-degree, 45-degree and 90-degree polarization images by using a polarization camera, and carrying out filtering and denoising processing on the polarization images by using deep learning so as to maximally reduce the interference of noise on a test; and S2, performing matching calibration calculation on the de-noised polarization image, obtaining a polarization degree image and a polarization angle image according to a formula, and inputting the polarization degree image and the polarization angle image into a deep learning training model as two channels. And S3, obtaining three features of a low layer, a middle layer and a high layer by using deep learning. And S4, increasing a receptive field by using deep learning to obtain an enhanced feature map, inputting the enhanced feature map into the SINet neural network, and searching to obtain a final wanted camouflage target feature from candidate features for matching in the image. And S5, for camouflage targets at different positions in different scenes, training by using the same neural network model, so that the model can still correctly identify the targets in different scenes and when the positions of the camouflage targets are different. The method can effectively identify the camouflage target hidden in the environment, and can be widely applied to the fields of multi-scene single or multi-target camouflage detection and the like.

Description

(I) Technical field

[0001] The present invention is a polarization imaging camouflage target recognition method based on deep learning. It utilizes polarization technology combined with a deep learning neural network to perform recognition operations on camouflage targets in different environments and positions, and belongs to the technical field of optical image information extraction. (II) Background technology

[0002] Polarization imaging has the advantage of detecting artificial targets based on their inherent characteristics. However, with the development of camouflage materials and camouflage shielding performance, the anti-optical detection technology of camouflaged targets has been continuously improved, and polarization imaging technology has been applied to many fields, including remote sensing, medical diagnosis, defogging, underwater imaging, face recognition, and target detection. In the field of target detection, previous studies have shown that polarization imaging technology can improve the contrast between artificial targets and natural backgrounds and suppress clutter signals of natural backgrounds due to the different material properties and surface roughness of artificial targets from natural backgrounds. However, the spatial resolution of infrared polarization imaging is currently insufficient to display the surface texture information of artificial targets. Therefore, we consider using passive polarization imaging technology in the visible light band to detect artificial targets camouflaged in natural scenes and obtain their surface texture information at the same time.

[0003] In order to improve the recognition ability of camouflaged targets, a polarization imaging camouflaged target recognition method based on deep learning is proposed. The traditional recognition ability of camouflaged targets using polarization is limited, and it takes a lot of manpower and material resources when facing a large number of images to be recognized; and some traditional deep learning methods only perform some preprocessing on polarization images, and cannot make better use of the features of all images. This paper proposes a new deep learning neural network model for camouflaged targets in different scenes and locations, that is, deep learning is introduced in the preprocessing stage and the feature extraction and fusion stage, and they are combined to better identify camouflaged targets. The recognition of camouflaged targets is achieved, which is more efficient than traditional polarization recognition and retains the background color of the original image; compared with traditional deep learning methods, the features recognized are more obvious and the effect is better. (III) Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying camouflaged targets using polarization imaging based on deep learning. This method solves the problem that the traditional polarization method cannot retain the background color of the original image and the traditional deep learning method cannot achieve better recognition results, greatly reduces the consumption of manpower and material resources, and realizes accurate recognition of camouflaged targets.

[0005] The object of the present invention is achieved in that:

[0006] A method for identifying camouflaged targets using polarization imaging based on deep learning comprises the following steps:

[0007] Step S1: Use a polarization camera to obtain 0°, 45°, and 90° polarization images, and use deep learning to filter and denoise the polarization images to minimize the interference of noise on the test.

[0008] Step S2: Perform matching calibration calculation on the denoised polarization image, obtain the polarization degree and polarization angle images according to the formula, and input them as two channels into the deep learning training model.

[0009] Step S3: Use deep learning to obtain low-level, mid-level, and high-level features.

[0010] Step S4: Use deep learning to increase the receptive field, obtain the enhanced feature map, input it into the SINet neural network, and search to obtain the final desired camouflaged target features from the candidate features. This part designs a new structure for matching in the image.

[0011] Step: S5: For camouflaged targets at different locations in different scenes, the same neural network model is used for training, so that the model can still correctly identify the target in different scenes and at different camouflaged target locations.

[0012] Preferably, the step S4: using deep learning to increase the receptive field, obtaining an enhanced feature map, inputting it into the SINet neural network, and searching, obtaining the final desired camouflaged target feature from the candidate features, this part adds a new structure:

[0013] Including: adding Bconv branches to maximize the receptive field, and setting a trainable kernel in each layer to selectively filter the features of each layer, retaining the layer features with good effects, so that the final feature map effect is optimal.

[0014] Preferably, the Bconv branch is added to maximize the receptive field, and a trainable kernel is set in each layer to selectively filter the features of each layer. The process of retaining the layer features with good effects is as follows:

[0015] Each branch performs a dilated convolution with a dilated number s as required, concatenates the feature maps of the first six branches, performs a 1×1 convolution dimensionality reduction operation on them, adds them to the seventh branch, and finally outputs the enhanced feature map. The polarization degree, polarization angle image, and the enhanced feature map in the original image are trained through SINet to filter out the required camouflaged target features, match them with the corresponding image, and finally fuse them and display them in the original image.

[0016] Preferably, during training, for camouflaged targets at different locations in different scenarios, the process of using the neural network model for training is as follows:

[0017] Before deep learning training, optimize the neural network hyperparameters and expand the existing data set to improve the model robustness and accuracy.

[0017] Preferably, the process of optimizing the neural network hyperparameters and expanding the existing data set is:

[0018] Replace the activation function and use the Mish activation function that is more suitable for the data set. Add a cosine annealing learning rate adjustment strategy to prevent the model from overfitting. Introduce more scene image data. When starting training, randomly split, paste, crop, change the angle of the data set to add more data sets and improve the robustness and accuracy of the model. (IV) Description of the drawings

[0019] Figure 1 Flowchart of the polarization imaging camouflage target recognition method based on deep learning. (V) Specific implementation methods

[0020] The present invention will be further described below in conjunction with specific embodiments.

[0021] In order to enable ordinary technicians in the field to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings. The specific process is as follows Figure 1 shown.

[0022] Step S1: Use a polarization camera to obtain 0°, 45°, and 90° polarization images, and use deep learning to filter and denoise the polarization images to minimize the interference of noise on the test.

[0023] The data acquisition model structure is composed of FD-1665 polarization imager (1, lens 2, 0°, 45°, 90° polarizers 3, 3 CCD sensors 4, three-way prism beam splitter). The training set and test set images collected by the polarization imager in the system are denoised by the RIDNet neural network model to obtain processed images.

[0024] Step S2: Perform matching calibration calculation on the denoised polarization image, obtain the polarization degree and polarization angle images according to the formula, and input them as two channels into the deep learning training model.

[0025] Through feature point matching, the three angle polarization images are aligned in space, and then the corresponding polarization degree and polarization angle images are obtained through matching calibration calculation.

[0026] Step S3: Use deep learning to obtain low-level, mid-level, and high-level features.

[0027] When the deep learning neural network model is working, the ResNet feature extractor obtains low-level image features through convolutional layers and pooling layers, obtains mid-level image features through residual blocks and their downsampling, and obtains high-level image features through deep residual blocks and global average pooling.

[0028] Step S4: Use deep learning to increase the receptive field, obtain the enhanced feature map, input it into the SINet neural network, and search to obtain the final desired camouflaged target features from the candidate features. This part designs a new structure for matching in the image.

[0029] Add a Bconv branch, perform a dilated convolution with a dilated number s on each branch as required, concatenate the feature maps of the first 6 branches, perform a 1×1 convolution dimensionality reduction operation on them, add them to the seventh branch, and finally output the enhanced feature map to increase the receptive field. After the enhanced feature map is obtained through the ResNet feature extractor, the processed training set images are input into the SINet neural network for training and the weight file is obtained.

[0030] Step: S5: For camouflaged targets at different locations in different scenes, the same neural network model is used for training, so that the model can still correctly identify the target in different scenes and at different camouflaged target locations.

[0031] Replace the activation function and use the Mish activation function that is more suitable for the data set. Add a cosine annealing learning rate adjustment strategy to prevent the model from overfitting. Introduce more scene image data. When starting training, randomly split, crop, and change the angle of the data set to add more data sets and improve the robustness and accuracy of the model.

[0032] This system can be used to identify camouflaged targets in grasslands, and can also be expanded to identify camouflaged targets in other environments and different camouflaged target categories. It can be widely used in the field of camouflaged target recognition.

[0033] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A polarization imaging camouflage target recognition method based on deep learning. Its characteristics are: The following steps are involved: Step S1: Use a polarization camera to obtain 0°, 45°, and 90° polarization images, and use deep learning to filter and denoise the polarization images to minimize the interference of noise on the test. Step S2: Perform matching calibration calculation on the denoised polarization image, obtain the polarization degree and polarization angle images according to the formula, and input them as two channels into the deep learning training model. Step S3: Use deep learning to obtain low-level, mid-level, and high-level features. Step S4: Use deep learning to increase the receptive field, obtain the enhanced feature map, input it into the SINet neural network, and search to obtain the final desired camouflaged target features from the candidate features. This part designs a new structure for matching in the image. Step: S5: For camouflaged targets at different locations in different scenes, the same neural network model is used for training, so that the model can still correctly identify the target in different scenes and at different camouflaged target locations.

2. The method according to claim 1, characterized in that: Deep learning is used to increase the receptive field, obtain the enhanced feature map, input it into the SINet neural network, and search to obtain the final desired camouflaged target features from the candidate features. A new structure is designed in this part, including: adding Bconv branches to maximize the receptive field, and setting a trainable kernel in each layer to selectively filter the features of each layer, retaining the layer features with good effects, so that the final feature map effect is the best.

3. The method according to claim 2, characterized in that: Add Bconv branches to maximize the receptive field, and set a trainable kernel in each layer to selectively filter the features of each layer, including: each branch performs a dilated convolution with a dilated number s as required, concatenates the feature maps of the first 6 branches, performs a 1×1 convolution dimensionality reduction operation on them, adds them to the seventh branch, and finally outputs the enhanced feature map. The polarization degree, polarization angle image, and the enhanced feature map in the original image are trained by SINet to filter out the required camouflaged target features, match them with the corresponding image, and finally fuse them and display them in the original image.

4. The method according to claim 1, characterized in that: The same neural network model is used for training camouflaged targets at different locations in different scenarios, so that the model can correctly identify targets in different scenarios and at different camouflaged target locations, including: optimizing the neural network hyperparameters and expanding the existing data set before deep learning training to improve the model robustness and accuracy.

5. The method according to claim 4, characterized in that: Optimize the neural network hyperparameters and expand the existing data set, including: changing the activation function, using the Mish activation function that is more suitable for the data set, adding a cosine annealing learning rate adjustment strategy to prevent model overfitting, introducing more scene image data, and randomly splitting, cropping, and changing the angle of the data set when starting training to add more data sets and improve the robustness and accuracy of the model.

Images