A low-resolution image multi-classification recognition method, device, equipment and medium

Abstract

The present application relates to the technical field of image recognition, and in particular to a low-resolution image multi-classification recognition method, device, equipment and medium, the method comprising the following steps: step 1: obtaining an original image containing a low-resolution target, and preprocessing the original image; the original image includes an algae image captured by high-speed photography; step 2: on the basis of a standard ResNet50 network, a channel-space dual attention mechanism and a feature fusion module are added to obtain an improved ResNet50-CBAM network, so as to classify and identify low-resolution algae in the original image; step 3: using an SGD optimizer, a dynamic update learning strategy is adopted to train the ResNet50-CBAM network; step 4: after the ResNet50-CBAM network is trained, the trained ResNet50-CBAM network is run to perform multi-classification recognition on the original image.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a method, apparatus, device, and medium for multi-classification recognition of low-resolution images. Background Technology

[0002] With the rapid development of computer vision technology, image classification plays a vital role in many fields, such as environmental monitoring, medical image analysis, and security surveillance. However, existing technologies still have significant shortcomings for low-resolution image classification tasks, especially in feature extraction, data adaptability, and computational efficiency, which face severe challenges, as detailed below:

[0003] Insufficient feature extraction capability: Traditional convolutional neural network (CNN) models (such as VGG, basic ResNet, etc.) are usually designed for high-resolution images, and their deep convolutional structures perform poorly on low-resolution images. Due to the sparse pixel information in low-resolution images, edge and texture features are significantly degraded, making it difficult for the model to effectively capture key discriminative features, resulting in a significant drop in classification accuracy.

[0004] Poor data adaptability: In real-world applications, low-resolution images often originate from various acquisition devices (such as high-speed cameras and surveillance cameras). Due to limitations in shooting conditions, these images frequently suffer from motion blur, uneven lighting, and noise interference. However, existing algorithms generally lack targeted preprocessing modules and fail to adaptively optimize for these degradation factors, further reducing the model's robustness and generalization ability.

[0005] Low computational efficiency: Some existing solutions employ a sequential processing flow of "super-resolution reconstruction + classification," which first improves image resolution using super-resolution technology before performing classification and recognition. While this approach can improve classification performance to some extent, it introduces additional computational overhead, leading to a significant increase in inference latency and making it difficult to meet the real-time requirements of applications such as industrial online monitoring and real-time video analysis.

[0006] Furthermore, low-resolution images present even greater challenges in multi-class object classification tasks. Due to the small size of the targets, the scarcity of feature information, and factors such as scale variations and background interference in the images, traditional detection algorithms struggle to achieve high-precision multi-class recognition under low-resolution conditions, severely limiting the practical application value of related technologies.

[0007] In summary, existing low-resolution image classification techniques have significant shortcomings in feature extraction, adaptive optimization, and computational efficiency. There is an urgent need for an efficient, robust, and lightweight solution to improve the classification accuracy of multi-class targets in low-resolution images while meeting real-time requirements. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this application provides a method, apparatus, device and medium for multi-class recognition of low-resolution images, aiming to improve the accuracy of multi-target classification in low-resolution images.

[0009] The technical means adopted by this invention to solve its technical problem is: a multi-classification recognition method, the improvement of which includes the following steps:

[0010] Step 1: Acquire the original image containing the low-resolution target and preprocess the original image; the original image includes algae images captured by a high-speed camera;

[0011] Step 2: Based on the standard ResNet50 network, a channel-spatial dual attention mechanism and a feature fusion module are added to obtain an improved ResNet50-CBAM network, which is used to classify and identify low-resolution algae in the original image.

[0012] Step 3: Train the ResNet50-CBAM network using the SGD optimizer and a dynamic update learning strategy;

[0013] Step 4: After the ResNet50-CBAM network has been trained, run the trained ResNet50-CBAM network to perform multi-class recognition on the original image.

[0014] The preprocessing of the original image described in the above technical solution includes:

[0015] Step 11: Apply random strategies to the original image, including but not limited to automatic contrast adjustment, tone separation, random flipping, color dithering, and solarization.

[0016] Step 12: Perform integrated adaptive illumination correction on the image after the random strategy combination, and perform local contrast limiting processing;

[0017] Step 13: After local contrast limiting processing, the image with size H×W is cropped to the target size S×S by scaling and random cropping.

[0018] Step 14: Construct a dynamic blur kernel based on high-speed camera parameters to simulate a real scene, and perform data augmentation on the scaled and cropped image;

[0019] Step 15: Perform mean normalization on the image after dynamic blurring.

[0020] Step 12 in the above technical solution includes the following steps:

[0021] Step 121: For a single channel, the given pixel value range is [0, L-1], and the total number of pixels is L;

[0022] Step 122: Crop the image into several sub-blocks of size M×N. Process each sub-block independently. Calculate the maximum number of pixels that each pixel can take in each restricted histogram according to formula (1).

[0023] (1);

[0024] Where β is the set contrast limit threshold, and h(i) represents the number of pixels with pixel value i in the sub-block, i = 0, 1, 2, ... L-1, N tile This represents the total number of pixels within the sub-block. ;

[0025] Step 123: Distribute the excess portion after cropping evenly across all possible pixel values:

[0026] (2);

[0027] in, , indicating the portion that exceeds the trimming length;

[0028] Step 124: Apply histogram equalization to each sub-block:

[0029] (3);

[0030] in, ;

[0031] Step 125: Perform bilinear interpolation on the equalization results of adjacent sub-blocks to calculate the final grayscale value:

[0032] (4);

[0033] Among them, s 11 s 12 s 21 s 22 The values ​​of w1, w2, w3, and w4 are the values ​​obtained after histogram equalization of the four adjacent sub-blocks of a pixel. w1, w2, w3, and w4 are weight coefficients based on pixel position.

[0034] Step 13 in the above technical solution includes the following steps:

[0035] Step 131: Randomly generate candidate cropping regions for each attempt. Calculate the area A of the target region respectively. target Randomly select the aspect ratio r and calculate the candidate width W. candidate High Hcandidate N attmpts This represents the maximum number of attempts.

[0036] (5);

[0037] Where Uniform(x,y) represents generating uniformly distributed random numbers in the range (x, y);

[0038] (6);

[0039] (7);

[0040] Among them, (r min ,r max () represents the aspect ratio range;

[0041] If W candidate ≤W and H candidate If the value is ≤H, proceed to the next step; otherwise, retry. If this step is the maximum number of attempts and the random width and height are still not successfully obtained, then perform center clipping.

[0042] Step 132: Randomly select the top left corner coordinate (X) of the image. min ,Y min ),satisfy The cropping area is The cropped area is then scaled to the target size S×S using bicubic interpolation.

[0043] Step 14 in the above technical solution includes the following steps:

[0044] Step 141: Construct a motion trajectory description in polar coordinates based on the angle parameter θ and the displacement length L, and calculate the displacement vector (dx, dy).

[0045] ,θ∈(-45°,45°)L∈(5,15)(8;

[0046] Step 142: Uniformly sample discrete points along the motion direction, calculate the cumulative coverage area of ​​each pixel position, and generate a normalized weight matrix MotionBlur by integrating the contribution of all sampling points on the path.

[0047] Step 143: Using a physics-driven gating coefficient α, dynamically adjust the fuzzy intensity based on the displacement length L to construct the output feature:

[0048] (9).

[0049] The momentum of the SGD optimizer described in the above technical solution is 0.9, the initial learning rate is 0.1×batch_size / 256, and the epoch decay is 30 / 60 / 90.

[0050] The improved ResNet50-CBAM network in step 2 of the above technical solution, which utilizes a channel-spatial dual attention mechanism and a feature fusion module, includes:

[0051] A CBAM layer is added after the last BottleNeck2 of each stage in the standard ResNet50 network;

[0052] Adjust the stride of the stage4 convolutional layer from 2 to 1, while keeping the output width at 14, the same as that of stage3.

[0053] A feature fusion module is embedded before global average pooling. The feature fusion module includes a concatenated layer, a two-dimensional convolutional layer with 2048 output channels, a batch normalization layer, and a ReLU activation layer.

[0054] The technical means adopted by this invention to solve its technical problem is: a low-resolution image multi-classification recognition device, used to implement the low-resolution image multi-classification recognition method as described above, the device comprising:

[0055] A data processing module is used to acquire raw images containing low-resolution targets and preprocess the raw images; the raw images include algae images captured by a high-speed camera.

[0056] Model building module: This module is used to improve the ResNet50-CBAM network by adding a channel-spatial dual attention mechanism and a feature fusion module to the standard ResNet50 network, so as to classify and identify low-resolution algae in the original image.

[0057] Model training module: used to train the ResNet50-CBAM network using the SGD optimizer and a dynamically updated learning strategy;

[0058] The classification and recognition module is used to run the trained ResNet50-CBAM network after the ResNet50-CBAM network is completed, and to perform multi-class recognition on the original image.

[0059] The means adopted by this invention to solve its technical problem is: a device comprising: at least one processor and at least one memory, wherein,

[0060] The memory stores program instructions or code;

[0061] The program instructions or code are loaded and executed by the processor, enabling the electronic device to implement the low-resolution image multi-class recognition method as described in any of the preceding claims.

[0062] The means adopted by the present invention to solve its technical problem is: a medium on which program instructions or code are stored, the program instructions or code being loaded and executed by a processor to realize the low-resolution image multi-class recognition method as described in any of the preceding claims.

[0063] The beneficial effects of this invention are:

[0064] ① By applying various data augmentation preprocessing methods, including motion blur and illumination correction (CLANE), an integrated dynamic data augmentation pipeline is formed that is specifically adapted to low-resolution high-speed motion images;

[0065] ② Applying the channel-space attention mechanism (CBAM), by combining the channel attention module and the spatial attention module, an improvement is made to the pre-trained model ResNet50, enabling the network to dynamically adjust the attention within the convolutional network;

[0066] ③ By combining the added feature fusion module and increasing the resolution of the ResNet50-CBAM output feature map, low-resolution feature enhancement can be achieved in multi-class recognition of low-resolution images (especially single-channel images captured by high-speed grayscale cameras). Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating a low-resolution image multi-class recognition method according to an embodiment of the present invention.

[0068] Figure 2 This is a flowchart illustrating step 1 of an embodiment of the present invention;

[0069] Figure 3 This is a flowchart illustrating step 12 in an embodiment of the present invention;

[0070] Figure 4 This is a flowchart illustrating step 13 in an embodiment of the present invention;

[0071] Figure 5 A flowchart illustrating step 14 of an embodiment of the present invention;

[0072] Figure 6 This is a schematic diagram of the ResNet50-CBAM network structure shown in an embodiment of the present invention;

[0073] Figure 7 This is a schematic diagram of the CBAM layer structure shown in an embodiment of the present invention;

[0074] Figure 8This is a structural block diagram of a low-resolution image multi-class recognition device according to an embodiment of the present invention;

[0075] Figure 9 This is a schematic diagram of the device shown in an embodiment of the present invention. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0077] The following will clearly and completely describe the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this invention can be combined interactively without contradicting each other.

[0078] like Figure 1 As shown, this application provides a low-resolution image multi-class recognition method, the method comprising:

[0079] Step 1: Acquire the original image containing the low-resolution target and preprocess the original image; the original image includes algae images captured by a high-speed camera.

[0080] As mentioned earlier, algae images captured by high-speed cameras suffer from problems such as motion blur and uneven illumination. Existing algorithms have not specifically optimized the preprocessing module. At the same time, some models have low computational efficiency. Due to their use of a cascaded super-resolution reconstruction scheme, inference latency is increased, which cannot meet the needs of real-time monitoring.

[0081] Based on this, this application provides a new preprocessing method, in one possible implementation, such as Figure 2 As shown, the preprocessing of the original image includes:

[0082] Step 11: Apply random strategies to the original image. These random strategies include, but are not limited to, automatic contrast adjustment, tone separation, random flipping, color dithering, and solarization.

[0083] In an 8-bit grayscale image, the brightness information of a single pixel can be represented by 0 (black) to 255 (white), with intermediate values ​​representing different shades of gray, thus describing the brightness of the pixel; the automatic contrast adjustment automatically maps the original image so that the darkest value becomes 0 and the brightest value becomes 225.

[0084] Tone separation involves setting the bits below a certain number of bits y in the binary value corresponding to pixel x in each color channel to 0. In an exemplary embodiment, the value range of variable y is [0, 4].

[0085] Random flipping involves flipping the original image at a random angle. In an exemplary embodiment, the flip angle is -30° to 30°.

[0086] Color jitter adjusts the brightness, contrast, and saturation of the original image, with an adjustment range of -40% to 40% in an exemplary embodiment.

[0087] The solarization operation involves changing the x-value of pixels above a certain threshold to 225-x.

[0088] This application combines two strategies randomly selected from the above two approaches for each original image to simulate complex interferences in real-world scenes, such as changes in lighting (e.g., overexposure / underexposure), differences in device imaging (e.g., color quantization error), and changes in viewing angle (e.g., tilted shooting angle), significantly expanding the distribution range of training data. By randomly selecting two strategies each time (e.g., automatic contrast + color jitter or tone separation + solarization), the model avoids memorizing fixed enhancement patterns and forces the model to learn more fundamental feature representations. This significantly improves the adaptability to the complexity of real-world scenes while maintaining inference efficiency, making it particularly suitable for scenarios with limited data or high annotation costs.

[0089] Step 12: Perform integrated adaptive illumination correction on the image after the random strategy combination, and perform local contrast limiting (CLAHE) processing.

[0090] Specifically, such as Figure 3 As shown, step 12 includes the following steps:

[0091] Step 121: For a single channel, the given pixel value range is [0, L-1], and the total number of pixels is L;

[0092] Step 122: Crop the image into several sub-blocks of size M×N. Process each sub-block independently. Calculate the maximum number of pixels that each pixel can take in each restricted histogram according to formula (1).

[0093] (1);

[0094] Where β is the set contrast limit threshold, h(i) represents the number of pixels with pixel value i in the sub-block (i = 0, 1, 2, ... L-1), and N tile This represents the total number of pixels within the sub-block. ;

[0095] Step 123: Distribute the excess portion after cropping evenly across all possible pixel values:

[0096] (2);

[0097] in, , indicating the portion that exceeds the trimming length;

[0098] Step 124: Apply histogram equalization to each sub-block:

[0099] (3);

[0100] in, ;

[0101] Step 125: Perform bilinear interpolation on the equalization results of adjacent sub-blocks to calculate the final grayscale value:

[0102] (4);

[0103] Among them, s 11 s 12 s 21 s 22 The values ​​of w1, w2, w3, and w4 are the values ​​obtained after histogram equalization of the four adjacent sub-blocks of a pixel. w1, w2, w3, and w4 are weight coefficients based on pixel position.

[0104] While randomized strategy combinations increase image diversity, prevent overfitting, and improve model robustness through data augmentation, these processes may still result in uneven lighting or localized contrast issues, requiring subsequent processing for correction.

[0105] The adaptive illumination correction provided in this embodiment can solve the illumination inconsistency problem caused by random strategy combinations (such as random flipping and color jitter) by dynamically adjusting the image brightness distribution, and repair the image quality degradation problem that may be introduced by random strategies, ultimately forming a high-quality input with high dynamic range, low noise, and significant features.

[0106] Step 13: After local contrast limiting processing, the image (size H×W) is cropped to the target size (S×S) by scaling and random cropping.

[0107] Specifically, such as Figure 4As shown, step 13 includes the following steps:

[0108] Step 131: Randomly generate candidate cropping regions for each attempt. Calculate the area A of the target region respectively. target Randomly select the aspect ratio r and calculate the candidate width W. candidate High H candidate N attmpts This represents the maximum number of attempts.

[0109] (5);

[0110] Where Uniform(x,y) represents generating uniformly distributed random numbers in the range (x, y);

[0111] (6);

[0112] (7);

[0113] Among them, (r min ,r max () represents the aspect ratio range;

[0114] If W candidate ≤W and H candidate If the value is ≤H, proceed to the next step; otherwise, retry. If this step is the maximum number of attempts and the random width and height are still not successfully obtained, then perform center clipping.

[0115] Step 132: Randomly select the top left corner coordinate (X) of the image. min ,Y min ),satisfy The cropping area is The cropped area is then scaled to the target size S×S using bicubic interpolation.

[0116] The above embodiments simulate the diversity of target proportions in real scenes by randomly selecting aspect ratios (e.g., 0.5-2.0) within the range of [rmin, rmax], forcing the model to learn proportion-invariant features; and randomly selecting the upper left corner coordinates (X_min, Y_min) to cover different regions of the image, avoiding the model from focusing too much on the central region (e.g., avoiding learning only "center offset" features), thus enhancing the ability to detect edge targets.

[0117] Meanwhile, a random cropping failure fallback mechanism is set up: when the random generation of candidate regions fails (such as exceeding the image boundary), the validity of the data is ensured by retrying or center cropping, avoiding training interruption caused by extreme image sizes. At the same time, center cropping serves as a backup strategy to preserve key content.

[0118] By combining dynamic scaling, multi-location sampling, and high-quality scaling strategies, the model's adaptability to target proportions, locations, and scales is systematically improved while ensuring data validity. At the same time, key details are preserved through anti-distortion interpolation, ultimately forming highly diverse, high-quality, and robust training data, which significantly improves the model's generalization performance in complex scenarios.

[0119] Step 14: Construct a dynamic blur kernel based on high-speed camera parameters to simulate a real scene, and perform data augmentation on the scaled and cropped image.

[0120] Since conventional data augmentation (such as fixed-angle rotation and symmetrical cropping) cannot simulate the dynamic blur and random motion trajectory of high-speed cameras, this application introduces a motion blur simulation layer. Based on high-speed camera parameters (such as exposure time and frame rate), a blur kernel simulating a real scene is dynamically generated. This dynamic blur kernel adopts a dynamic parameterized convolution kernel generation technique.

[0121] Specifically, such as Figure 5 As shown, step 14 includes the following steps:

[0122] Step 141: Construct a motion trajectory description in polar coordinates based on the angle parameter θ and the displacement length L, and calculate the displacement vector (dx, dy).

[0123] ,θ∈(-45°,45°)L∈(5,15)(8;

[0124] Step 142: Uniformly sample discrete points along the motion direction, calculate the cumulative coverage area of ​​each pixel position, and generate a normalized weight matrix MotionBlur by integrating the contribution of all sampling points on the path.

[0125] Step 143: Using a physics-driven gating coefficient α, dynamically adjust the fuzzy intensity based on the displacement length L to construct the output feature:

[0126] (9).

[0127] The above embodiments, through a three-tiered linkage design of parametric motion modeling, physical blur generation, and adaptive fusion, achieve accurate simulation and intelligent suppression of motion blur while maintaining computational efficiency. Its core value lies in constructing an end-to-end learnable framework from motion representation to feature enhancement, significantly improving the model's feature extraction capabilities and robustness in dynamic scenes.

[0128] Step 15: Perform mean normalization on the image after dynamic blurring.

[0129] Specifically, mean = [123.675, 116.28, 103.53],

[0130] std=[58.395, 57.12, 57.375],

[0131] Norm(x) = (x - mean) / std;

[0132] Where x is the pixel value of the input image; mean represents the mean vector of each channel, corresponding to the three RGB channels; and std represents the standard deviation vector of each channel.

[0133] Step 2: Based on the standard ResNet50 network, a channel-spatial dual attention mechanism and a feature fusion module are added to obtain an improved ResNet50-CBAM network, which is used to classify and identify low-resolution algae in the original image.

[0134] To address the issue that the standard ResNet50 network has an output width of 7 in stage 4, resulting in severe loss of spatial information in low-resolution images, this application makes the following improvements to the standard ResNet50 network:

[0135] Specifically, such as Figure 6 As shown, a CBAM layer is added after the last BottleNeck2 of each stage in the standard ResNet50 network;

[0136] Adjust the stride of the stage4 convolutional layer from 2 to 1, while keeping the output width at 14, the same as that of stage3.

[0137] A feature fusion module is embedded before global average pooling, and the main body of the feature fusion module is... Figure 6 The diagram is outlined in dashed lines and includes a concatenation layer (Concatenate()), a 2D convolutional layer with 2048 output channels, a batch normalization layer, and a ReLU activation layer, thereby aggregating the features of stage3 (14×14) and stage4 (14×14) to enhance the ability to recognize small targets.

[0138] The improved ResNet50-CBAM network provided in this application adds a channel-spatial dual attention mechanism (referencing the CBAM improvement) after the last BottleNeck2 of each stage in the standard ResNet50 stages 1 to 4. This mechanism prioritizes the activation of high-frequency signals in algae edge textures. The channel-spatial dual attention mechanism can dynamically enhance the attention in deep convolutional networks like ResNet50-CBAM.

[0139] On the one hand, CBAM can consider the global dependencies between each channel in the feature map in the channel dimension, automatically identify important feature channels, and suppress noisy channels; on the other hand, it can simultaneously consider the dependencies of each spatial location in the feature map relative to its neighboring locations, thereby focusing on the target region, weakening background interference, and improving the overall performance of deep convolutional networks. A schematic diagram of the CBAM layer structure in this embodiment is shown below. Figure 7 As shown.

[0140] Step 3: Train the ResNet50-CBAM network using the SGD optimizer and a dynamic update learning strategy.

[0141] In one exemplary embodiment, the momentum of the SGD optimizer is 0.9, the initial learning rate is 0.1×batch_size / 256, and the epoch decay is 30 / 60 / 90.

[0142] Step 4: After the ResNet50-CBAM network has been trained, run the trained ResNet50-CBAM network to perform multi-class recognition on the original image.

[0143] Through the above embodiments, this application innovatively applies a variety of data augmentation preprocessing methods, including motion blur and illumination correction (CLANE), to form an integrated dynamic data augmentation pipeline specifically adapted to low-resolution high-speed motion images;

[0144] Simultaneously, the channel-space attention mechanism (CBAM) is applied. By combining the channel attention module and the spatial attention module, the network is improved on the basis of the pre-trained model ResNet50, enabling the network to dynamically adjust the attention within the convolutional network.

[0145] Finally, by combining the added feature fusion module and increasing the resolution of the ResNet50-CBAM output feature map, low-resolution feature enhancement is achieved in multi-class recognition of low-resolution images (especially single-channel images captured by high-speed grayscale cameras).

[0146] The following are embodiments of the apparatus described in this application, which can be used to execute the low-resolution image multi-class recognition method involved in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments of the low-resolution image multi-class recognition method involved in this application.

[0147] Please see Figure 8 This application provides a low-resolution image multi-class recognition device 50, which includes: a data processing module 501, a model building module 502, a model training module 503, and a classification and recognition module 504.

[0148] The data processing module 501 is used to acquire an original image containing a low-resolution target and to preprocess the original image; the original image includes an algae image captured by a high-speed camera.

[0149] Model building module 502: This module is used to add a channel-spatial dual attention mechanism and a feature fusion module to the standard ResNet50 network to obtain an improved ResNet50-CBAM network, which is used to classify and identify low-resolution algae in the original image.

[0150] Model training module 503: used to train the ResNet50-CBAM network using the SGD optimizer and a dynamically updated learning strategy;

[0151] The classification and recognition module 504 is used to run the trained ResNet50-CBAM network after the ResNet50-CBAM network is completed, and to perform multi-class recognition on the original image.

[0152] It should be noted that the low-resolution image multi-class recognition device provided in the above embodiments is only illustrated by the division of the above functional modules when performing low-resolution image multi-class recognition. In practical applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the low-resolution image multi-class recognition device will be divided into different functional modules to complete all or part of the functions described above. The above modules can be embedded in hardware or independent of the processor in the computer device, or they can be stored in software in the memory of the computer device so that the processor can call and execute the operations corresponding to the above modules.

[0153] Furthermore, the low-resolution image multi-class recognition device and the low-resolution image multi-class recognition method provided in the above embodiments belong to the same concept. The specific way in which each module performs its operation has been described in detail in the method embodiments, and will not be repeated here.

[0154] Please see Figure 9 This application provides an electronic device 4000.

[0155] exist Figure 9In this design, data interaction between the processor 4001 and the memory 4003 can be achieved through at least one communication bus 4002. This communication bus 4002 may include a path for transmitting data between the processor 4001 and the memory 4003. The communication bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 4002 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0156] Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of this application.

[0157] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc. The memory 4003 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program instructions or code in the form of instructions or data structures and accessible by the electronic device 400, but not limited thereto.

[0158] The memory 4003 stores program instructions or code, and the processor 4001 can read the program instructions or code stored in the memory 4003 through the communication bus 4002.

[0159] When the program instructions or code are executed by the processor 4001, the low-resolution image multi-class recognition method in the above embodiments is implemented.

[0160] Furthermore, this application provides a storage medium storing program instructions or code, which is loaded and executed by a processor to implement the low-resolution image multi-class recognition method described above.

[0161] This application provides a computer program product, which includes program instructions or code. The program instructions or code are stored in a storage medium. The processor of an electronic device reads the program instructions or code from the storage medium, loads and executes the program instructions or code, so that the electronic device implements the low-resolution image multi-class recognition method as described above.

[0162] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for multi-class recognition of low-resolution images, characterized in that, The method includes: Step 1: Acquire the original image containing the low-resolution target and preprocess the original image; the original image includes algae images captured by a high-speed camera; specifically, the preprocessing of the original image includes: Step 11: Apply random strategies to the original image, including but not limited to automatic contrast adjustment, tone separation, random flipping, color dithering, and solarization. Step 12: Perform integrated adaptive illumination correction on the image after combining the random strategies, and perform local contrast limiting processing; further, step 12 includes the following steps: Step 121: For a single channel, the given pixel value range is [0, L-1], and the total number of pixels is L; Step 122: Crop the image into several sub-blocks of size M×N. Process each sub-block independently. Calculate the maximum number of pixels that each pixel can take in each restricted histogram according to formula (1). （1）； Where β is the set contrast limit threshold, and h(i) represents the number of pixels with pixel value i in the sub-block, i = 0, 1, 2, ... L-1, N tile This represents the total number of pixels within the sub-block. ; Step 123: Distribute the excess portion after cropping evenly across all possible pixel values: （2）； in, , indicating the portion that exceeds the trimming length; Step 124: Apply histogram equalization to each sub-block: （3）； in, ; Step 125: Perform bilinear interpolation on the equalization results of adjacent sub-blocks to calculate the final grayscale value: （4）； Among them, s 11 s 12 s 21 s 22 The values ​​of w1, w2, w3, and w4 are the weight coefficients based on pixel position, obtained by independently performing histogram equalization on the four adjacent sub-blocks of a certain pixel. Step 13: After local contrast limiting processing, the image with size H×W is cropped to the target size S×S by scaling and random cropping. Step 14: Construct a dynamic blur kernel based on high-speed camera parameters to simulate a real scene, and perform data augmentation on the scaled and cropped image; Step 14 includes the following steps: Step 141: Construct a motion trajectory description in polar coordinates based on the angle parameter θ and the displacement length L, and calculate the displacement vector (dx, dy). ,θ∈(-45°,45°),L∈(5,15)(8); Step 142: Uniformly sample discrete points along the motion direction, calculate the cumulative coverage area of ​​each pixel position, and generate a normalized weight matrix MotionBlur by integrating the contribution of all sampling points on the path. Step 143: Using a physics-driven gating coefficient α, dynamically adjust the fuzzy intensity based on the displacement length L to construct the output feature: （9）； Step 15: Perform mean normalization on the image after dynamic blurring; Step 2: Based on the standard ResNet50 network, a channel-spatial dual attention mechanism and a feature fusion module are added to obtain an improved ResNet50-CBAM network, which is used to classify and identify low-resolution algae in the original image. Step 3: Train the ResNet50-CBAM network using the SGD optimizer and a dynamic update learning strategy; Step 4: After the ResNet50-CBAM network has been trained, run the trained ResNet50-CBAM network to perform multi-class recognition on the original image.

2. The low-resolution image multi-class recognition method according to claim 1, characterized in that, Step 13 includes the following steps: Step 131: Randomly generate candidate cropping regions for each attempt. Calculate the area A of the target region respectively. target Randomly select the aspect ratio r and calculate the candidate width W. candidate High H candidate N attmpts This represents the maximum number of attempts. （5）； Where Uniform(x,y) represents generating uniformly distributed random numbers in the range (x, y); （6）； （7）； Among them, (r min ,r max () represents the aspect ratio range; If W candidate ≤W and H candidate If the value is ≤H, proceed to the next step; otherwise, retry. If this step is the maximum number of attempts and the random width and height are still not successfully obtained, then perform center clipping. Step 132: Randomly select the top left corner coordinate (X) of the image. min ,Y min ),satisfy The cropping area is The cropped area is then scaled to the target size S×S using bicubic interpolation.

3. The low-resolution image multi-class recognition method according to claim 1, characterized in that, The SGD optimizer has a momentum of 0.9, an initial learning rate of 0.1×batch_size / 256, and epoch decay of 30 / 60 / 90.

4. The low-resolution image multi-class recognition method according to claim 1, characterized in that, Step 2 includes: A CBAM layer is added after the last BottleNeck2 of each stage in the standard ResNet50 network; Adjust the stride of the stage4 convolutional layer from 2 to 1, while keeping the output width at 14, the same as that of stage3. A feature fusion module is embedded before global average pooling. The feature fusion module includes a concatenated layer, a two-dimensional convolutional layer with 2048 output channels, a batch normalization layer, and a ReLU activation layer.

5. A low-resolution image multi-class recognition device, used to implement the low-resolution image multi-class recognition method according to claim 1, characterized in that, include A data processing module is used to acquire raw images containing low-resolution targets and preprocess the raw images; the raw images include algae images captured by a high-speed camera. Model building module: This module is used to improve the ResNet50-CBAM network by adding a channel-spatial dual attention mechanism and a feature fusion module to the standard ResNet50 network, so as to classify and identify low-resolution algae in the original image. Model training module: used to train the ResNet50-CBAM network using the SGD optimizer and a dynamically updated learning strategy; The classification and recognition module is used to run the trained ResNet50-CBAM network after the ResNet50-CBAM network is completed, and to perform multi-class recognition on the original image.

6. A device, characterized in that, include: At least one processor, at least one memory, wherein, The memory stores program instructions or code; The program instructions or code are loaded and executed by the processor, causing the electronic device to implement the low-resolution image multi-class recognition method as described in any one of claims 1 to 4.

7. A medium having program instructions or code stored thereon, characterized in that, The program instructions or code are loaded and executed by the processor to implement the low-resolution image multi-class recognition method as described in any one of claims 1 to 4.

Images