An ammunition target image recognition platform based on clarification processing

Abstract

This invention discloses a munition target image recognition platform based on sharpening preprocessing, comprising: a high-performance learning and training platform for sharpening preprocessing raw munition images to construct a dataset, and training a multi-scale feature fusion network based on the dataset, outputting a lightweight inference model through pruning and INT8 quantization; and an integrated intelligent detection terminal integrating an image acquisition module and an embedded inference subsystem; the image acquisition module is connected to the embedded inference subsystem for real-time acquisition of battlefield munition target images; the embedded inference subsystem is used to deploy the lightweight inference model, achieving pipelined parallel processing of image acquisition and model inference through DMA acceleration and a double buffering mechanism. This invention, through offline pre-processing sharpening enhancement technology, enables the front-end terminal to achieve efficient recognition with short single-image response time and high embedded recognition accuracy even without complex preprocessing.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recognition technology, specifically relating to a munition target image recognition platform based on sharpening processing, which is suitable for real-time recognition of battlefield munition targets by airborne platforms such as UAVs. Background Technology

[0002] With the widespread use of the internet and social media, obtaining image information of missile targets from public channels has become an important intelligence tool. However, these images are usually taken by non-professionals and are subject to limitations such as ambient lighting, high-speed movement, and network transmission compression, resulting in serious degradation, blurring, and inconsistent resolution, making it difficult to identify key target features.

[0003] Existing image target classification methods face challenges when dealing with specific military targets, including complex features, high similarity between categories, limitations of multi-scale and small sample sizes, and the conflict between real-time performance and hardware deployment. These shortcomings result in low efficiency for automated identification of high-value military targets in uncontrolled environments, making it difficult to meet the requirements of accuracy, robustness, and deployment flexibility for practical applications. Therefore, there is an urgent need for an automatic missile type identification technology that can achieve high precision, lightweight design, and embedded deployment even under low-quality image conditions. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides a munition target image recognition platform based on sharpening processing, including a high-performance learning and training platform and an integrated intelligent detection terminal.

[0005] The technical solution for achieving the objective of this invention is as follows:

[0006] A platform for recognizing ammunition targets based on sharpening processing, comprising:

[0007] A high-performance learning and training platform is used to perform sharpening preprocessing on raw ammunition images to build a dataset, and to train a multi-scale feature fusion network based on the dataset. A lightweight inference model is output through pruning and INT8 quantization.

[0008] The integrated intelligent detection terminal integrates an image acquisition module and an embedded inference subsystem;

[0009] The image acquisition module is connected to the embedded inference subsystem and is used to acquire images of battlefield munition targets in real time.

[0010] The embedded inference subsystem is used to deploy the lightweight inference model and achieves pipelined parallel processing of image acquisition and model inference through DMA acceleration and double buffering mechanism.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] (1) By performing Gaussian filtering, sharpening and gamma correction on the offline training platform, the model can complete the robust learning of degradation features such as blur and dark light during the training stage, thus allowing the front-end integrated terminal to maintain accurate identification of munition targets with complex backgrounds without omitting complex preprocessing.

[0013] (2) Integrating the OV2640 sensor and the STM32F407 microcontroller into the same terminal shortens the physical path of the photoelectric signal to the processing unit. With the DMA transfer mode, the delay caused by image preprocessing and cross-module communication in the traditional solution is eliminated, and the single image recognition response time is ≤0.5 seconds.

[0014] (3) In the offline stage, a comprehensive evaluation index combining MSE and SSIM is adopted. While enhancing the edge details of the image, it effectively suppresses the interference of artifacts and noise on the training set, so that the recognition accuracy of the model reaches 86.5% in the small sample ammunition dataset.

[0015] (4) The model size was compressed from 22MB to 2.5MB through pruning and INT8 quantization, and directly embedded into the kernel in a hard-coded form. By eliminating the complex inference engine layer and front-end preprocessing layer, the Flash usage of the system was significantly reduced, and the recognition accuracy remained above 80% in the embedded environment.

[0016] (5) The integrated terminal design greatly reduces the hardware size and system power consumption, and can be quickly deployed as an independent sensing payload on airborne platforms such as UAVs and loitering munitions where space is limited and real-time requirements are extremely high. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall framework of the simulation training platform of the present invention;

[0018] Figure 2 This is a flowchart of the image preprocessing process of the present invention;

[0019] Figure 3 This is a flowchart to clarify the evaluation process of this invention.

[0020] Figure 4 This is a flowchart of the model training process of the present invention;

[0021] Figure 5 , 6 Figures 7 and 8 show the hardware interface circuit and connection diagrams of the embedded platform of this invention.

[0022] Figure 9 This is a physical diagram showing the connection of the modules of this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below in conjunction with specific embodiments.

[0024] The target image recognition platform based on sharpening processing disclosed in this invention has an overall architecture including a high-performance learning and training platform and an integrated intelligent detection terminal, as detailed below:

[0025] 1. High-performance learning and training platform

[0026] The high-performance learning and training platform employs an offline computing model training subsystem. Its hardware core is a GPU cluster, used to train a multi-scale feature fusion network based on a munitions target dataset. Backpropagation is used to optimize weights to suppress overfitting on small samples, and high-precision offline testing is performed before deployment. The overall framework of the model training subsystem is as follows: Figure 1 As shown, the system first reads parameters from the configuration file and preprocesses the dataset images. Then, it instantiates the network model and optimizer by importing the pre-files and loss function. Next, it enters the core training and testing loop. During this period, the training data is saved in real time, and logical judgment is made based on the test accuracy. If the accuracy meets the target, the final model and results are saved; otherwise, the current training statistics are output and displayed directly.

[0027] The configuration file will pre-set all the parameters needed for the subsequent experiments, such as the learning rate, learning rate decay coefficient, and maximum training batch size. After training is complete, the data storage address also needs to be set here for easy retrieval and storage later.

[0028] The ammunition target dataset is a self-built multi-scale small-sample dataset, covering 25 representative aerodynamic ammunition targets. The samples are mainly derived from typical images selected from publicly available sources. For each target class, 10-20 core training samples and 5 algorithm test samples are obtained, totaling over 300 images. This multi-scale small-sample dataset is used to verify the robustness and generalization performance of the classification algorithm. To ensure the system accurately extracts core geometric features such as the projectile's aspect ratio, warhead curvature, and tail fin layout, the dataset selection primarily uses side-view images, supplemented by some front-side images, to enhance the model's ability to analyze three-dimensional spatial attitude. Considering the complexity of real-world combat conditions, the samples include different paint colors for the same target and introduce extreme cases such as cluttered backgrounds, interference from multiple targets, and incomplete target images, thus ensuring the trained system has strong environmental adaptability. Due to environmental and equipment factors, the clarity of the acquired images may not be guaranteed; therefore, image clarity needs to be improved before formal target classification and recognition. The image preprocessing flow of this invention is as follows: Figure 2 As shown.

[0029] The specific steps for image sharpening are as follows:

[0030] Step 1: Noise Reduction. Gaussian filtering is used to remove Gaussian noise from the image. Gaussian filtering is a linear smoothing filter that replaces the current pixel value with the weighted average of its neighboring pixels after performing a convolution operation on the image. Gaussian filtering is suitable for removing Gaussian noise and smoothing the image. It mainly consists of two steps:

[0031] (1) Generate a Gaussian kernel. Specific process: First, construct an empty matrix based on the coordinates of each position. According to the requirement, the size of the matrix should be an odd number greater than 3, usually 3 or 5. Take the center point of the matrix as the origin. Taking a 5×5 matrix as an example, the coordinates of the upper left corner are (-2, 2); Second, calculate the weight corresponding to each coordinate using a two-dimensional Gaussian function; Finally, assign the calculated weight values ​​to the corresponding positions of the coordinate matrix to obtain the weight matrix and normalize it. The weight of the corresponding coordinate is represented by the following formula:

[0032]

[0033] In the formula, (x, y) are the coordinates of a point within the Gaussian kernel relative to the geometric center of the Gaussian kernel. The standard deviation of the two-dimensional Gaussian function itself is changed. It can change the numerical distribution within the Gaussian kernel, thereby controlling the image processing results. The larger the value, the more dispersed the weights, resulting in stronger blurring and noise reduction effects, but it is also easier to lose some details.

[0034] The weight matrix used in this study is 3×3, with a standard deviation of [missing information]. Take 1, Setting the value to 1 effectively smooths out subtle high-frequency noise points in the image without making the image too blurry, thus preserving the complete structure of the image. This yields the weight matrix:

[0035]

[0036] After obtaining the weight matrix, normalization is required to avoid affecting the overall brightness of the image. This normalization process yields the necessary Gaussian kernel.

[0037]

[0038] (2) After obtaining the weight matrix, traverse the entire image. Here, a 3×3 Gaussian kernel is used. The first pixel is selected as the second pixel in the second row and second column as the starting point. Then, select the surrounding 3×3 pixels to obtain the pixel matrix at that point. Multiply the weight matrix with the pixel matrix at that point one by one, and then... The new pixel value for that point is then assigned. The matrix shows that the center point has a higher weight, consistent with the natural law that the closer the point, the greater its influence.

[0039] Suppose an image contains noise, where the pixel value of that point is significantly different from the surrounding pixels:

[0040]

[0041] After processing, the pixel value of this point is 107.4, and the point becomes noticeably smoother than its surroundings.

[0042] For the complete image, repeat the above operation (2) until the entire image is traversed, and finally obtain a brand new image with noise reduction.

[0043] Step 2, Image Sharpening. This step enhances the edge and detail information of an image, improving its contrast and sharpness. Its core objective is to enhance high-frequency information (such as edges and contours) in the image, making the boundaries of objects in the image more distinct. Image sharpening involves the following steps:

[0044] (1) Define a sharpening convolution kernel, which is usually a small 3×3 matrix. The value at the geometric center of the matrix is ​​generally a large positive number, while the values ​​around it are negative. This is to enhance the difference between the center pixel and its neighboring pixels. The four-neighborhood sharpening kernel is a commonly used sharpening convolution kernel. It can enhance edges in both horizontal and vertical directions and introduces less noise. The four-neighborhood sharpening kernel is as follows:

[0045]

[0046] (2) Perform convolution operation, sliding the defined sharpening convolution kernel pixel by pixel on the image from left to right and from top to bottom. For each pixel in the image, perform the following operations: align the center of the convolution kernel with the current pixel, multiply each value in the convolution kernel with the corresponding pixel value on the image, add all the product results, and the sum is the new value of the current pixel after sharpening.

[0047] (3) Repeat the convolution operation of step (2) above for each pixel in the image to finally generate a complete sharpened image.

[0048] Step 3: Apply gamma correction to enhance contrast. This is a non-linear operation that improves the overall visual effect of the image by adjusting the brightness values ​​of pixels. Its simplest form is defined by the power law formula:

[0049]

[0050] It is the input normalized pixel value. `γ` is the normalized pixel value of the output, `c` is the correction value (a positive real number), and `c` is a constant that changes the brightness of the image. When `c` is less than 1, the overall image becomes darker; when `c` is greater than 1, the overall image becomes brighter. To ensure the image satisfies the constraint of mapping from an interval to itself, it is usually set to 1. Gamma correction with `c` set to 1 / 2.2 is performed before image storage, and inverse correction with `c` set to 2.2 is performed during display. The effect of the gamma value `γ` on image brightness and contrast is as follows: When `γ=1`, the image remains unchanged. If the overall image or region of interest is dark, setting `0≤γ<1` can increase image contrast. If the overall image or region of interest is bright, setting `γ>1` can decrease image contrast.

[0051] Step 4: Grayscale Conversion. For subsequent object classification and recognition, the color image can be converted to grayscale mode to increase the training load and improve learning performance. The grayscale image is obtained using a weighted calculation formula:

[0052]

[0053] The image uses an RGB color space model, where red, green, and blue colors are displayed with different intensity weights, corresponding to R, G, and B in the formula, respectively.

[0054] Step 5: Size Normalization. Since images from different sources vary in size, their dimensions need to be standardized. A background-filling method is used: images of different sizes are first scaled to a specified uniform size, then the scaled image is placed in the center and the background is filled in, thus achieving image size uniformity. This method avoids feature extraction failures caused by direct stretching or cropping.

[0055] A common method for evaluating image sharpness is gradient-based assessment. This method uses the Laplacian operator to extract image feature edges and calculates the variance to evaluate the sharpness effect. While this method is fast and has clear physical meaning, it has a chance of misinterpreting noise points as image edge details due to the evaluation mechanism, leading to an inflated sharpness score. To address this, the images used in this experiment were all obtained after sharpening. Therefore, a combined evaluation metric is used to compare the images before and after processing. The evaluation combines the mean squared error (MSE) and structural similarity (SSIM) of the processed image with those of the original image.

[0056] MSE calculates the expected value of the square of the difference between corresponding gray values ​​of two images. The smaller the value, the less information is lost in the processed image. Its calculation formula is as follows:

[0057]

[0058] In the formula, M and N are the length and width dimensions of the image, I(i,j) is the gray value of the original image at pixel coordinates (i,j), and K(i,j) is the gray value of the sharpened image at pixel coordinates (i,j). The smaller the MSE, the more original features the image retains.

[0059] SSIM, on the other hand, comprehensively considers the brightness, contrast, and structure of an image. This method can compensate for the inconsistency between MSE and human visual perception, as the selected criteria are parameters that are highly sensitive to the human eye. The calculation method is as follows:

[0060]

[0061] In the formula, M and N are the image dimensions. , Let x be the mean of the original image x and the processed image y at (i, j) (a measure of brightness). , Let x be the variance (a measure of contrast) of the original image x and the processed image y at (i, j). Let x be the covariance of the original image x and the processed image y at (i, j) (a measure of structural correlation). It is a constant for maintaining stability (to prevent the denominator from being 0), where Take 6.5025, We set the value to 58.5225. The final range of SSIM is (-1, 1), and the closer it is to 1, the more the image highlights the features of the original image.

[0062] The commonly used method combining both evaluation criteria is a hybrid loss function strategy. However, the optimal weights of different parameters change during training, leading to local optima in later stages. To ensure image sharpness and platform response speed, this paper proposes multiplying the sensitivity of MSE to error and SSIM to structure to achieve mutual constraint. The final evaluation formula is as follows:

[0063]

[0064] The smaller the MSE, the higher the degree of image feature preservation. Meanwhile, the SSIM value ranges from -1 to 1, therefore, the clearer the image, the better. The closer the score is to 0, the less image distortion there is. This method considers both the image's gradient and various parameters, thus preserving image integrity very well. The evaluation process is as follows: Figure 3 .

[0065] The multi-scale feature fusion network used in this invention is MScNet10, with the following structure: 5 convolutional layers, 4 fully connected layers, a feature map size of 4×6, 255 feature maps, and approximately 56 million parameters. The specific training and testing process of the model is as follows: Figure 4 As shown, an adaptive learning rate is used. After each batch of training, the loss function value is recorded, and the model is tested, with the accuracy recorded. The model is updated when the test results are better. After the model training is complete, it needs to be compressed through pruning and INT8 quantization to output a lightweight model file, improving the model's running efficiency as much as possible without affecting the running performance.

[0066] Pruning removes redundant parameters generated during training, such as some low-contribution convolutional kernels or feature channels. These parameters have little impact on image recognition accuracy but consume a significant amount of computation time, hence the need for pruning to eliminate them. The core of pruning lies in evaluating the importance of the parameter structure and setting or removing unimportant parts, thereby obtaining a more convenient network model. During pruning, iterative processing is performed based on one or more structures, typically selecting weight magnitudes, gradients, and image features. In this training example, convolutional kernels and feature channels were selected as iteration objects. Then, the L1 norm was calculated to represent the importance index of the convolutional kernel, and importance was ranked. Five training epochs were then performed. The reason for performing only a few training epochs is to prevent pruning from affecting the overall accuracy. After training, pruning compresses the model, outputting a lightweight model file for faster subsequent image recognition processing.

[0067] The specific method for implementing INT8 convolutional inference on STM32F4 is as follows: The input feature map is first multiplied by the input transformation scaling factor (scale) to convert it to int8 type. The calculation formula is as follows:

[0068]

[0069] In the formula To determine the maximum absolute value of the data to be quantized in floating-point format, and to ensure that all floating-point numbers can be mapped to the INT8 range without overflowing, and since the INT8 range is [-127, 127], 127 is used as the upper limit.

[0070] The 3×3 convolution is looped and unrolled, then the SMLAD multiply-accumulate instruction is used to accumulate the convolution result into an int32 variable. The accumulated result is multiplied by the weights. The weights are the product of the input scale. Mathematical Definitions and Inputs Similar to scale, the calculation formula is as follows:

[0071]

[0072] In the formula, X refers to the set of all floating-point values ​​in the feature map.

[0073] After ReLU activation, the output is an int8 value. Fully connected layers also use INT8 quantization. Model weights are stored in external SPI Flash (e.g., W25Q64, 64Mbit capacity), loaded into a designated area of ​​internal SRAM via DMA, and read layer by layer during inference.

[0074] 2. Integrated intelligent detection terminal

[0075] The integrated intelligent detection terminal includes an image acquisition module and an embedded platform.

[0076] The image acquisition module uses an OV2640 image sensor, connected to the embedded inference subsystem (embedded platform) via a DCMI interface. It acquires real-time images of battlefield munitions and targets, outputting RGB24 format data (each pixel consists of three 8-bit channels: red, green, and blue, totaling 24 bits). The OV2640 sensor's configuration parameters are: maximum resolution 1632×1232, transmission speed 54MB / s, and module size 32mm×21mm. This invention sets its output resolution to 1280×720 and the frame rate to 30fps. The sensor registers are initialized via the SCCB bus, including setting the output window, pixel clock polarity, and frame rate.

[0077] The embedded platform is based on the STM32F407 microcontroller, with built-in FPU and DSP instruction sets, a main frequency of 168MHz, 1MB of program memory, and 256K×8 of RAM. The camera serves directly as the image acquisition component of the embedded platform. Signal interconnection with the main MCU is achieved via its I / O ports.

[0078] Image data is directly transferred from the OV2640 sensor to the STM32F407 memory using DMA technology, without occupying the CPU during the transfer process. A double-buffering mechanism is employed to enable parallel execution of image acquisition, sharpening processing, and model inference pipelines. The specific configuration is as follows: the DMA channel is configured with the source address being the DCMI data register and the destination addresses being two buffers allocated in SRAM. An interrupt is triggered after each image frame is acquired, notifying the CPU to begin processing the current buffer, while the DMA continues to transfer the next image frame to the other buffer. This double-buffered pipeline effectively reduces end-to-end latency; experiments show that the response time for single-image recognition is ≤0.5 seconds.

[0079] The hardware for DMA hardware acceleration technology includes: a DMA controller, an image acquisition interface, a double-buffered memory, an interrupt line, and a CPU core. The input of the DMA controller is connected to the image acquisition interface, the output is connected to the double-buffered memory, and the interrupt line is connected to the CPU core. In cyclic acquisition mode, the DMA controller automatically transfers image data alternately to the two buffers of the double-buffered memory. When either buffer is full, an interrupt is triggered to notify the CPU core to perform inference, while data continues to be transferred to the other buffer, forming a pipelined parallel processing.

[0080] Specific wiring methods and interface circuit designs are as follows: Figure 5 , 6 As shown in Figures 7 and 8.

[0081] Figure 5 The demonstration showcases a power management system designed for image sensors. It uses two low-dropout linear regulators to convert the input 3.3V voltage into two independent voltages and utilizes an RC filter circuit to provide clean power to the analog section.

[0082] The upper part of the circuit is built based on the first regulator U1 (PAM3101DAB28), with a fixed output voltage of 2.8V. The second capacitor C2 (10µF) and the third capacitor C3 (1µF) are connected in parallel between the input voltage 3V3 and GND. The second capacitor (C2) stores charge to handle transient current demands, and the third capacitor (C3) filters out high-frequency noise to ensure stable input voltage. The 2.8V voltage from the OUT pin of the first regulator U1 is initially filtered by the fifth capacitor C5, and then shunted to two different networks: the first analog power supply (AVDD) and the second analog power supply (DOVDD). A first capacitor C1 is connected to AGND (analog ground) at the AVDD terminal of the first analog power supply. The first resistor R1 and the first capacitor C1 form an RC low-pass filter connected between the fifth capacitor C5 and AGND (analog ground). The first resistor impedes current surges, and the first capacitor filters out high-frequency noise to prevent digital noise from interfering with image quality. At the second analog power supply DOVDD terminal, there are two parallel capacitors, the tenth capacitor C10 (4.7µF) and the eleventh capacitor C11 (0.1µF), connected to GND. The parallel capacitors are connected in series with the second resistor R2 and then connected to the fifth capacitor C5. RC filtering is also used, but here two capacitors with different capacitance values ​​are used to filter out low-frequency and high-frequency noise at the same time, ensuring the stability of digital signal transmission.

[0083] The lower half of the circuit is built based on the second regulator U2 (PAM3101DAB120). The input pins VIN and EN are connected to 3V3, providing a fixed output voltage of 1.2V to power the core logic of the chip. The GND pin is grounded. The sixth capacitor C6 acts as an input decoupling capacitor, connected between 3V3 and GND. The eighth capacitor C8 (10µF) and the ninth capacitor C9 (0.1µF) are connected in parallel between the positive power supply terminal VCC and the ground terminal GND. The eighth capacitor C8 stabilizes the voltage, and the ninth capacitor C9 filters out high-frequency noise, ensuring stable operation of the core logic circuit. The seventh capacitor C7 is connected in series between the BTT pin and ground GND. The third resistor R3 is a 0-ohm resistor. Using a 0-ohm resistor to connect GND (digital ground) and AGND (analog ground) ensures voltage consistency and utilizes the impedance characteristics of the resistor to prevent high-frequency noise from the digital ground from entering the analog ground, protecting the purity of the analog signal.

[0084] Figure 6 The circuit on the left is configured for the clock input, with two input sources, XCLK1 and XCLK2, providing two different clock sources. Resistors R4 and R5 are two 0-ohm resistors (jump resistors) that act as a "switch" or "selector." When resistor R4 is soldered, the XCLK1 signal is transmitted through it to the XCLK network, and then input to camera pin 13. When resistor R5 is soldered, the XCLK2 signal is transmitted through it.

[0085] The J1 connector on the right is the physical interface between the main control board and the camera module. The J1 connector interfaces are divided into three main categories: power, control, and data.

[0086] 1. Power supply pins

[0087] Pin 1 is connected to analog ground (AGND) to reduce noise interference.

[0088] Pin 3 (AVDD) is connected to the analog power supply to power the analog part of the camera.

[0089] Pin 9 (DVDD) is connected to the digital core power supply, which powers the internal digital logic of the sensor.

[0090] Pin 10 (DOVDD) is connected to the digital I / O power supply to power the digital interface.

[0091] Pin 15 (DGND) is connected to digital ground.

[0092] 2. Control Interface

[0093] Pin 4 (SIO_C) connects to the SCCB clock line and is used to configure the serial clock input of the camera.

[0094] Pin 5 (RESET) is the reset pin. It is activated by either a low or high input level to restart the camera sensor, depending on the requirements.

[0095] Pin 6 (VSYNC) is connected to the vertical synchronization signal, indicating the start of a frame of image data.

[0096] Pin 7 (PWDN) is the power-down / sleep mode control, used to control the camera to enter low-power mode.

[0097] Pin 8 (HREF) is connected to the horizontal reference signal (horizontal synchronization), indicating the effective transmission period of one line of image data.

[0098] Pin 13 (XCLK) is connected to the system master clock input.

[0099] Pin 16 (PCLK) is the pixel clock output. It is the clock signal that accompanies the camera's output data, with each PCLK corresponding to one pixel data.

[0100] 3. Data output interface

[0101] Pin 2 (SIO_D) connects to the SCCB data line. It is used to configure the serial data input / output of the camera's internal registers.

[0102] Pins 11-12, 14, 16-24 (D0-D9) are image data output pins. This is a parallel data bus, through which the camera transmits the acquired image pixel data to the main control chip.

[0103] Figure 7 This is a pinout diagram of the OV2640 camera sensor. The chip in the middle of the diagram represents the OV2640 image sensor chip, which supports image acquisition up to 2 megapixels (1632x1232).

[0104] The OV2640 has 20 pins, which are divided into three types: power, control, and data. The specific connections are as follows:

[0105] 1. Power supply pins

[0106] Pin 7 is grounded.

[0107] Pin 9 is connected to the chip's main power input, which is 3.3V.

[0108] 2. Control Interface

[0109] Pin 1 is a hardware reset pin, used to restart the sensor.

[0110] Pin 2 is connected to the VSYNC vertical sync signal, indicating the start of a frame of image.

[0111] Pin 3 is the clock line for the SCCB interface, used to synchronize the transmission of configuration data.

[0112] Pin 4 is connected to PWDN to control the camera to enter a low-power state.

[0113] Pin 5 is the data line for the SCCB interface, used for bidirectional transmission of configuration commands and parameters.

[0114] Pin 6 is connected to the HREF horizontal reference signal, indicating the start of a line of valid image data.

[0115] Pin 10 is connected to the system master clock of the XCLK1 external input to provide the operating clock source for the sensor.

[0116] Pin 16 is connected to the PCLK pixel clock signal, and each clock pulse corresponds to the output of one pixel data.

[0117] 3. Data output interface

[0118] Pins 8, 11-15, and 17-20 connect to D0-D9, a 10-bit parallel data bus used to output image pixel data. In practical applications, an 8-bit data mode (D0-D7) is typically used, in which case D8 and D9 may be left floating or used for other functions.

[0119] Figure 8 This is a schematic diagram of an active crystal oscillator circuit, which provides a 12MHz clock signal to the system.

[0120] Figure 9 This is an image captured when the microcontroller is working after all pins are connected.

[0121] Although specific embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A platform for recognizing ammunition targets based on sharpening processing, characterized in that, include: A high-performance learning and training platform is used to perform sharpening preprocessing on raw ammunition images to build a dataset, and to train a multi-scale feature fusion network based on the dataset. A lightweight inference model is output through pruning and INT8 quantization. The integrated intelligent detection terminal integrates an image acquisition module and an embedded inference subsystem; The image acquisition module is connected to the embedded inference subsystem and is used to acquire images of battlefield munition targets in real time. The embedded inference subsystem is used to deploy the lightweight inference model and achieves pipelined parallel processing of image acquisition and model inference through DMA acceleration and double buffering mechanism.

2. The ammunition target image recognition platform based on sharpening processing according to claim 1, characterized in that, The sharpening preprocessing process includes: Perform Gaussian filtering for noise reduction: Generate a Gaussian kernel, which is a weight matrix calculated from a two-dimensional Gaussian function. After obtaining the weight matrix, multiply the weight matrix with the pixel matrix of each pixel to obtain the new pixel value. Repeat this process for the entire image to finally obtain a brand new image with noise reduction. Image sharpening: Define a sharpening convolution kernel, perform convolution operations, slide the defined sharpening convolution kernel pixel by pixel on the image from left to right and from top to bottom. For each pixel in the image, perform the following operations: align the center of the sharpening convolution kernel with the current pixel, multiply each value in the sharpening convolution kernel with the pixel value at the corresponding position in the image, add all the product results, and the sum is the new value of the current pixel after sharpening, finally generating a complete sharpened image; Gamma correction enhances contrast: It improves the overall visual effect of an image by adjusting the brightness value of pixels; Grayscale conversion: Converting a color image to grayscale mode; Size normalization: Images of different sizes are first scaled to a specified size according to a uniform standard, and then the scaled image is placed in the center of the image and the background is filled to achieve image size uniformity.

3. The ammunition target image recognition platform based on sharpening processing according to claim 2, characterized in that, During the high-performance learning and training platform phase, a hybrid evaluation metric combining mean squared error (MSE) and structural similarity index (SSIM) is used to assess the sharpening effect. The final evaluation formula is as follows: When the final evaluation score S is closer to 0, the image distortion is smaller and the feature retention is higher.

4. The ammunition target image recognition platform based on sharpening processing according to claim 2, characterized in that, Gamma correction enhances contrast according to the power law formula: ； It is the input normalized pixel value. γ is the normalized pixel value of the output, c is the correction value, and c is a constant that changes the brightness of the image. Before storing the image, gamma correction with c set to 1 / 2.2 is performed, and inverse correction with c set to 2.2 is performed during display. Let 0≤γ<1 to increase image contrast; let γ>1 to decrease image contrast.

5. The ammunition target image recognition platform based on sharpening processing according to claim 1, characterized in that, The multi-scale feature fusion network consists of 5 convolutional layers and 4 fully connected layers, with a feature map size of 4×6, 255 feature maps, and approximately 56 million parameters.

6. The ammunition target image recognition platform based on sharpening processing according to claim 1, characterized in that, The INT8 quantization maps floating-point data to the range [-127, 127] by calculating the input conversion scaling factor (scale). The calculation formula is as follows: ; In the formula This represents the maximum absolute value of the data to be quantized in floating-point format. The 3×3 convolution is looped and unrolled, then the SMLAD multiply-accumulate instruction is used to accumulate the convolution result into an int32 variable; the accumulated result is then multiplied by the weights. The weights are the product of the input transformation scaling factor (scale). The calculation formula is as follows: ； In the formula, X refers to the set of all floating-point values ​​in the feature map; The ReLU activation function outputs an INT8 result.

7. The ammunition target image recognition platform based on sharpening processing according to claim 1, characterized in that, The dual-buffering mechanism configures the DMA channel, uses the DCMI data register as the source address, and uses two regions allocated in SRAM as target buffers. When an interrupt is triggered after a frame of image acquisition is completed, the CPU processes the data in the current buffer, and the DMA synchronously transfers the next frame of image to the other buffer.

8. The ammunition target image recognition platform based on sharpening processing according to claim 1, characterized in that, The power management system of the image acquisition module uses two low-dropout linear regulators to convert the input 3.3V voltage into two independent voltages. The first path includes a first regulator U1, a first capacitor C1, a second capacitor C2, a third capacitor C3, a fifth capacitor C5, a first resistor R1, a second resistor R2, and a tenth capacitor C10. The second capacitor C2 and the third capacitor C3 are connected in parallel between the input voltage 3V3 and GND. The second capacitor C2 is used to store charge to meet transient current demands, and the third capacitor C3 is used to filter out high-frequency noise. The 2.8V voltage from the OUT pin of the first regulator U1 is initially filtered by the fifth capacitor C5, and then split to the first analog power supply and the second analog power supply. The first capacitor C1 is connected to AGND. The first resistor R1 and the first capacitor C1 form an RC low-pass filter, which is connected between the fifth capacitor C5 and AGND. The second analog power supply has two parallel tenth capacitors C10 and eleventh capacitor C11 connected to GND. The parallel tenth capacitor is connected in series with the second resistor R2 and then connected to the fifth capacitor C5. The first path includes a second voltage regulator U2, a sixth capacitor C6, a seventh capacitor C7, an eighth capacitor C8, a ninth capacitor C9, and a third resistor R3. The input pins VIN and EN of the second voltage regulator U2 are connected to 3V3, fixing the output voltage to 1.2V, and the GND pin is grounded. The sixth capacitor C6 is connected between 3V3 and GND as an input decoupling capacitor. The eighth capacitor C8 and the ninth capacitor C9 are connected in parallel between the positive terminal VCC and the ground terminal GND. The eighth capacitor C8 stabilizes the voltage, and the ninth capacitor C9 filters out high-frequency noise. The seventh capacitor C7 is connected in series between the BTT pin and ground GND. The third resistor R3 is a 0-ohm resistor, connecting GND and AGND together.

Images