AI-based infrared target automatic identification method, equipment and medium

Abstract

The invention relates to the field of target detection and recognition, and discloses an AI-based infrared target automatic recognition method and device and a medium, and the method comprises the steps: collecting infrared image data containing wild animal targets, and building an initial infrared image data set; preprocessing the image of the data set to obtain a preprocessed data set; data expansion is carried out on the data set after image preprocessing, the sample size of the data set is increased, the updated data is labeled, target labeling information is obtained, and a final data set is obtained; and constructing an improved ALSS-YOLO infrared target recognition model, and training and optimizing the improved ALSS-YOLO infrared target recognition model by using the final data set to obtain an optimal target recognition model. According to the method, the field target can be effectively identified, and the method has relatively high identification precision and real-time performance.

Description

Technical Field

[0001] The present invention relates to the field of target detection and recognition, and in particular to an AI-based automatic infrared target recognition method. Background Art

[0002] Traditional infrared target recognition methods, both domestically and internationally, generally employ statistical pattern recognition methods based on feature fusion and matching, which are highly effective. However, these algorithms are ineffective in detecting and recognizing infrared targets in the wild, where their motion characteristics change rapidly or the target is very small. Deep learning algorithms and architecture design for image target recognition and detection have been a focus of recent research in computer vision. Convolutional neural networks do not require manual feature set. Deep learning algorithms can directly extract the required features from images using pre-designed models during training to perform target recognition. This approach improves feature quality and significantly enhances recognition accuracy. Summary of the Invention

[0003] The purpose of the present invention is to propose an AI-based infrared target automatic recognition method, device and medium to solve the technical problem that the current traditional infrared target recognition is not suitable for high-speed moving objects and small target objects.

[0004] Specifically, the present invention provides an AI-based infrared target automatic recognition method, comprising the following steps:

[0005] S1. Collect infrared image data containing wild animal targets and establish an initial infrared image dataset;

[0006] S2. Preprocess the images of the data set to obtain a preprocessed data set;

[0007] S3, perform data expansion on the dataset after image preprocessing, increase the sample size of the dataset, annotate the updated data, obtain target annotation information, and obtain the final dataset;

[0008] S4. Build an improved ALSS-YOLO infrared target recognition model and use the final dataset to train and optimize it to obtain the optimal target recognition model.

[0009] S5. Utilize the optimal target recognition model to perform target recognition and obtain target location information.

[0010] A storage medium storing instructions and data for implementing an AI-based infrared target automatic recognition method.

[0011] An AI-based infrared target automatic recognition device includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement an AI-based infrared target automatic recognition method.

[0012] The beneficial effects provided by the present invention are:

[0013] The AI-based automatic infrared target recognition method proposed in this paper significantly improves recognition accuracy compared to traditional methods through image preprocessing, dataset expansion, and optimization based on an improved ALSS-YO10 deep learning network. This method effectively solves the problem of target recognition in infrared images. This method can effectively identify targets in the wild with high recognition accuracy and real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a simple flow chart of the method of the present invention;

[0015] Figure 2 It is a schematic diagram of the working of the hardware device of an embodiment of the present invention. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0017] Before formally explaining the present invention, the scheme of the present invention is first generally explained for easy understanding.

[0018] Please refer to Figure 1 The present invention provides an AI-based infrared target automatic recognition method, comprising:

[0019] S1. Collect infrared image data containing wild animal targets and establish an initial infrared image dataset;

[0020] It should be noted that in step S1 of the present invention, an infrared uncooled thermal imager is used to collect infrared images of different animal targets in a field environment, at different weather conditions, at different temperatures, and at different times. The collected data is cleaned, data with abnormal imaging is removed, and an initial data set is established.

[0021] S2. Preprocess the images of the data set to obtain a preprocessed data set;

[0022] It should be noted that the preprocessing in step S2 includes non-uniformity correction, image noise reduction, contrast enhancement and detail enhancement.

[0023] Specifically, for the image after non-uniformity correction, the NLbayes-based denoising algorithm is used to perform edge-preserving denoising on the infrared image.

[0024] The contrast enhancement based on neighborhood information statistics is applied to the denoised image, as shown in formula (1):

[0025]

[0026] Among them, r is the neighborhood radius; λ r =(2r+1) 2 is the total number of pixels in the neighborhood, t is the threshold, is a binarization function, as shown in Formula 2:

[0027]

[0028] Combined with detail enhancement algorithm, the contrast of infrared images and the recognizability of infrared targets are improved.

[0029] S3, perform data expansion on the dataset after image preprocessing, increase the sample size of the dataset, annotate the updated data, obtain target annotation information, and obtain the final dataset;

[0030] Step S3 mainly involves expanding the dataset by rotating, scaling, translating, and flipping the infrared images after image preprocessing to increase data diversity. The objects appearing in the infrared images in the final dataset are annotated to obtain their location and category information.

[0031] S4. Build an improved ALSS-YOLO infrared target recognition model and use the final dataset to train and optimize it to obtain the optimal target recognition model.

[0032] In step S4, the labeled infrared dataset is used for model training to obtain an infrared target recognition model for identifying target information in infrared images. In view of the problems of low accuracy and slow speed of current target recognition networks, the present invention improves the improved ALSS-YO10 network to achieve the purpose of adapting to infrared images.

[0033] 1) Introducing the densely connected network (DenseNet), using the feedforward mode to connect each layer of information to other layers, that is, the current layer receives all the feature maps of the previous n-1 layers as input, as shown in formula (3):

[0034] x n =H n (x0,x1,...,x n-1 )(3)

[0035] Among them, x0,x1,...,x n-1 It is the concatenation of the feature maps of the first n-1 layers, and H represents the function that processes the cascaded features. Introducing a densely connected network can reduce the feature information loss caused by convolution and downsampling and increase feature transfer.

[0036] 2) Improve Batch Normalization. When training the neural network, normalize each hidden layer, not just when inputting the training set. This can improve the training speed and accuracy of the w and b parameters. The calculation is as follows:

[0037] Calculate the mean of the n-layer output, as shown in Formula 4

[0038] m=mean(z i )(4)

[0039] Calculate the variance of the n-layer output as shown in Formula 5:

[0040]

[0041] Normalized output, as shown in Formula 6:

[0042]

[0043] Among them, z represents the i-th parameter of the n-th layer output, ξ is the regularization coefficient to prevent the denominator from being zero, and s represents the number of output parameters of the n-th layer.

[0044] 3) Use CIoU Loss as the loss function to improve the positioning accuracy of the target, as shown in formula (7).

[0045]

[0046] where b and b gt represents the center point of the two rectangular boxes, ρ represents the Euclidean distance between the two rectangular boxes, and c represents the distance between the diagonals of the closure area of the two rectangular boxes.

[0047] Where v is used to measure the consistency of the relative proportions of the two rectangular boxes, and α is the weight coefficient;

[0048]

[0049] The CIoU loss is defined as:

[0050]

[0051] An infrared target recognition algorithm model is constructed with the improved ALSS-YOLO as the basic framework, and the infrared target recognition algorithm model is trained and optimized using the training set and validation set of the sample data set, and the model indicators are tested using the test set to obtain the optimal target recognition model.

[0052] S5. Utilize the optimal target recognition model to perform target recognition and obtain target location information.

[0053] In step S5, the optimal target recognition model is used to identify the infrared image of the wild animal target terminal to obtain the location of the wild animal.

[0054] See Figure 2 , Figure 2 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically includes: an AI-based infrared target automatic recognition device 401, a processor 402 and a storage medium 403.

[0055] An AI-based infrared target automatic recognition device 401: The AI-based infrared target automatic recognition device 401 implements the AI-based infrared target automatic recognition method.

[0056] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the AI-based infrared target automatic recognition method.

[0057] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the AI-based infrared target automatic recognition method.

[0058] The present invention provides an embodiment of infrared target recognition of wild animals in a nature reserve, which is specifically as follows:

[0059] 1. Data Collection and Initial Dataset Construction

[0060] Equipment and scenes: Infrared thermal imaging equipment is used to collect infrared images in different weather conditions (sunny, cloudy, rainy), temperatures (-10℃ to 30℃), and time periods (day and night) in the nature reserve. Target animals include deer, wolves, hares, etc.

[0061] Data cleaning: Images with blurred or excessive noise due to device jitter, occlusion, or extreme environments were eliminated, and 2,000 valid infrared images were finally selected to construct the initial data set.

[0062] 2. Image Preprocessing

[0063] Non-uniformity correction: A two-point correction method is used to eliminate image non-uniformity caused by inconsistent detector response.

[0064] Edge-preserving denoising: The NLBayes algorithm is applied, with the search window size set to 7×7, the similarity window to 3×3, and the noise standard deviation σ = 0.1, preserving the target edge details.

[0065] Contrast enhancement: The neighborhood information statistics method is used, with a neighborhood radius of r = 5, a threshold of t = 0.3, and a binarization function calculated according to formula (2) to enhance the contrast between the target and the background.

[0066] Detail enhancement: Combined with the CLAHE algorithm (Contrast-Limited Adaptive Histogram Equalization), the block size is 8×8 and the contrast limit is 2.0.

[0067] 3. Dataset expansion and annotation

[0068] Data augmentation: The preprocessed images were randomly rotated (±15°), scaled (0.8-1.2 times), horizontally flipped, and translated (±10% pixels) to expand to 8000 images.

[0069] Labeling: Use the LabelImg tool to label the target location (Bounding Box) and category (deer, wolf, hare) and generate a labeling file in XML format.

[0070] 4. Model training and optimization

[0071] Network architecture: Based on the improved ALSS-YOLO model, the backbone network introduces DenseNet-121, and the dense block connection method is implemented according to formula (3). Each layer receives the feature map input of the previous 4 layers.

[0072] Batch Normalization improvement: Add a BN layer after each convolution layer, calculate according to formulas (4)-(6), set the regularization coefficient ε=1e-5, and the sliding average momentum β=0.9.

[0073] Loss function: CIoU Loss (Formula 7-9) is used with a weight coefficient α = 0.6 to optimize target positioning accuracy.

[0074] Training parameters: Use Adam optimizer, initial learning rate 0.001, batch size 16, train for 50 epochs, and the validation set accuracy reaches 98.2%.

[0075] 5. Target recognition test

[0076] Test environment: 10 infrared videos (5 minutes each) were shot at night in a protected area (ambient temperature -5°C) and input into the trained model for real-time recognition.

[0077] Results: The model successfully identified 23 deer, 5 wolves, and 15 hares, with an average positioning accuracy (IoU) of 92.7% and a single-frame processing time of ≤30ms, meeting real-time requirements.

[0078] The beneficial effects of the present invention are:

[0079] The AI-based automatic infrared target recognition method proposed in this paper significantly improves recognition accuracy compared to traditional methods through image preprocessing, dataset expansion, and optimization based on an improved ALSS-YO10 deep learning network. This method effectively solves the problem of target recognition in infrared images. This method can effectively identify targets in the wild with high recognition accuracy and real-time performance.

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An AI-based infrared target automatic recognition method, characterized by: The following steps are involved: S1. Collect infrared image data containing wild animal targets and establish an initial infrared image dataset; S2. Preprocess the images of the data set to obtain a preprocessed data set; S3, perform data expansion on the dataset after image preprocessing, increase the sample size of the dataset, annotate the updated data, obtain target annotation information, and obtain the final dataset; S4. Build an improved ALSS-YOLO infrared target recognition model and use the final dataset to train and optimize it to obtain the optimal target recognition model; S5. Utilize the optimal target recognition model to perform target recognition and obtain target location information.

2. The AI-based infrared target automatic recognition method according to claim 1, characterized in that: The preprocessing in step S2 includes non-uniformity correction, image noise reduction, contrast enhancement and detail enhancement.

3. The AI-based infrared target automatic recognition method according to claim 2, characterized in that: The formula for contrast enhancement in step S2 is as follows: Among them, r is the neighborhood radius; λ r =(2r+1) 2 is the total number of pixels in the neighborhood, t is the preset threshold, x(h,w) represents the pixel value at coordinates h,w, and x(hk,wl) represents a neighborhood of size k×l centered at (h,w); is a binarization function, as shown in formula (2):

4. The AI-based infrared target automatic recognition method according to claim 1, wherein: The data expansion in step S3 includes rotation, scaling and translation.

5. The AI-based infrared target automatic recognition method according to claim 1, characterized in that: The improved ALSS-YOLO infrared target recognition model in step S4 is as follows: A densely connected network is introduced into the traditional ALSS-YOLO network, and the feedforward mode is used to connect each layer of information to other layers; When training the network, normalize each hidden layer; CIoU Loss is used as the loss function.

6. The AI-based infrared target automatic recognition method according to claim 5, characterized in that: The calculation formula of CIoULoss is as follows: where b and b gt represents the center point of the two rectangular boxes, ρ represents the Euclidean distance between the two rectangular boxes, and c represents the distance between the diagonals of the closure area of the two rectangular boxes; Where v is used to measure the consistency of the relative proportions of the two rectangular boxes, and α is the weight coefficient.

7. A storage medium, characterized in that: The storage medium stores instructions and data for implementing the AI-based infrared target automatic recognition method described in any one of claims 1 to 6.

8. An AI-based infrared target automatic recognition device, characterized by: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement the AI-based infrared target automatic recognition method described in any one of claims 1 to 6.

Images