Image recognition method and device, storage medium and electronic equipment

Abstract

The invention provides an image recognition method and device, a storage medium and electronic equipment, and solves the problems of poor low-illumination image enhancement effect and low recognition accuracy in the prior art. The method comprises the following steps: judging and classifying acquired images, dividing the acquired images into a first image set and a low-illumination image set, and enhancing the low-illumination image into a second image set; inputting the first image set and the second image set into an improved deep learning model for target detection, and training the model; and the to-be-recognized image is input into the trained model to form a recognition result, so that the enhancement effect of the low-illumination image is remarkably improved, and the recognition accuracy of the improved model is further improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an image recognition method, device, storage medium and electronic device. Background Art

[0002] In many imaging scenarios, the image quality is not ideal due to shooting reasons such as light and reflection from the object surface. The low-light environment poses a severe challenge to image recognition tasks.

[0003] Traditional methods, such as histogram equalization, improve contrast by stretching the image pixel value distribution, but they easily introduce noise, which damages image details, especially in dark areas, where details are difficult to effectively restore.

[0004] With the development of deep learning, methods based on convolutional neural networks (CNNs) have gradually emerged. These methods leverage the network's powerful feature learning capabilities to extract more discriminative features from low-light images for recognition. However, existing CNN architectures suffer from gradient fading in low-light scenarios and fail to fully consider the impact of uneven illumination on feature extraction, making it impossible to accurately locate target areas.

[0005] In addition, most current image recognition methods have poor adaptability to different types of low-light images (such as nighttime dim light, indoor low light, etc.), and it is difficult to maintain stable and efficient recognition performance in different scenarios, which seriously restricts their large-scale promotion and application in practical application fields such as security monitoring, autonomous driving, medical imaging, and power system operation and maintenance.

[0006] In summary, existing image recognition technology has shortcomings in image enhancement effect, feature extraction accuracy, and cross-scene adaptability. There is an urgent need for an image recognition method to meet the needs of practical applications. Summary of the Invention

[0007] The present invention provides an image recognition method, device, storage medium and electronic equipment to solve the problems of poor low-illumination image enhancement effect and low recognition accuracy in the prior art.

[0008] In a first aspect, the present invention provides an image recognition method, comprising the following steps: Acquiring images, determining the quality of the images, and forming a first image set (i.e., a set of images that do not require image enhancement, and are non-low-light images) and a low-light image set; performing enhancement processing on the low-illumination image set to form a second image set; Constructing a deep learning model, and training the deep learning model using the first image set and the second image set to form a trained deep learning model; Obtain the image to be identified and judge the quality of the image to be identified. If no enhancement is required, input the image to be identified into the trained deep learning model. Otherwise, input the enhanced image to be identified into the trained deep learning model to form a recognition result.

[0009] Preferably, the image quality is judged by one or more features of the image; wherein the features include CDF (Cumulative Distribution Function), HSV (Hue, Saturation, Value) and LAB (CIELAB, color space).

[0010] Preferably, acquiring images, judging the quality of the images, and forming a first image set and a low-illumination image set specifically include the following steps: Acquire an image and generate CDF features, HSV features, and LAB features of the image; Performing feature fusion on the CDF feature, the HSV feature, and the LAB feature to form a first fused feature; Inputting the first fusion features into the SVM linear kernel, the SVM polynomial kernel and the SVM Gaussian kernel respectively to form a first classification result, a second classification result and a third classification result respectively; The first classification result, the second classification result, and the third classification result are processed by an Adaboost enhanced classifier to form a first image set and a low-illumination image set.

[0011] Preferably, performing enhancement processing on the low-illumination image set to form a second image set specifically comprises the following steps: Converting the first RGB image of the low-light image set into a first color space HSV; Smoothing V in the first color space HSV by using a mean filter to form a first brightness component; Processing the first luminance component through guided filtering to form a second luminance component; performing logarithmic calculation on the second brightness component to form a reflection component; Optimizing the reflection component through a Gabor filter to form a third brightness component; Combining the second luminance component and the third luminance component to form a component V_enhanced; The color space is reorganized according to the component V_enhanced to form a second color space (H, S, V_enhanced), and the second color space is converted into a second RGB image (ie, a second image set).

[0012] Preferably, the deep learning model is a Yolo series model.

[0013] Preferably, the deep learning model is a Yolo v11n model; wherein the Yolo v11n model includes a backbone network (Baackbone), a connection layer (Neck), and an output layer (Head) connected in sequence; The backbone network sequentially connects the first convolutional (Conv) layer, the second convolutional layer, the first C3K2 module, the third convolutional layer, the second C3K2 module, the fourth convolutional layer, the third C3K2 module, the fifth convolutional layer, the fourth C3K2 module, the SPFF module and the C2PSA module; The connection layer comprises a first connection subnetwork and a second connection subnetwork, wherein the first connection subnetwork includes a first UpSample (up sampling) module, a first Concat (splicing) module, a fifth C3K2 module, a second UpSample module, a second Concat module, and a sixth C3K2 module connected in sequence; and the second connection subnetwork includes a sixth convolutional layer, a third Concat module, a seventh C3K2 module, a seventh convolutional layer, a fourth Concat module, and an eighth C3K2 module connected in sequence; The fifth C3K2 module is connected to the third Concat module, and the sixth C3K2 module is connected to the sixth convolutional layer; The output layer includes a first Detect module, a second Detect module and a third Detect module, which do not affect each other; The second C3K2 module is connected to the second Concat module, the third C3K2 module is connected to the first Concat module, the C2PSA module is connected to the first UpSample module and the fourth Concat module, the sixth C3K2 module is connected to the sixth convolutional layer and the first Detect module, the seventh C3K2 module is connected to the second Detect module, and the eighth C3K2 module is connected to the third Detect module.

[0014] Preferably, the Bottleneck module in the C3K2 module is replaced with a lightweight StarBlock module to form an I-C3K2 module.

[0015] Preferably, the first Detect module, the second Detect module and the third Detect module each include two parallel lines, one including two convolutional layers Conv, a two-dimensional convolutional layer Conv2d and a bounding box loss CLoU connected in sequence, and the other including a depth-separable convolutional layer DW-Conv, a convolutional layer Conv, a depth-separable convolutional layer DW-Conv, a convolutional layer C8onv, a two-dimensional convolutional layer Conv2d and a classification loss CLSLoss connected in sequence.

[0016] Preferably, one or more of the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the seventh convolutional layer are replaced with the corresponding first depth convolutional layer, the second depth convolutional layer, the third depth convolutional layer, the fourth depth convolutional layer, the fifth depth convolutional layer, the sixth depth convolutional layer and the seventh depth convolutional layer, and / or one or more C3K2 modules in the first C3K2 module, the second C3K2 module, the third C3K2 module, the fourth C3K2 module, the fifth C3K2 module, the sixth C3K2 module, the seventh C3K2 module and the eighth C3K2 module are replaced with the corresponding first I-C3K2 module, the second I-C3K2 module, the third I-C3K2 module, the fourth I-C3K2 module, the fifth I-C3K2 module, the sixth I-C3K2 module, the seventh I-C3K2 module and the eighth I-C3K2 module, respectively, to form an improved Yolo v11 model.

[0017] In a second aspect, the present invention further provides an image recognition device, which specifically includes the following modules: An image quality judgment module is used to acquire images, judge the quality of the images, and form a first image set (i.e., a set of images that do not require image enhancement) and a set of low-light images to be enhanced; a low-illumination image enhancement module, configured to perform enhancement processing on the low-illumination image set to form a second image set; A model training module, configured to construct a deep learning model and train the deep learning model using the first image set and the second image set to form a trained deep learning model; The recognition module is used to obtain the image to be recognized and judge the quality of the image to be recognized. If no enhancement is required, the image to be recognized is input into the trained deep learning model. Otherwise, the enhanced image to be recognized is input into the trained deep learning model to form a recognition result.

[0018] Preferably, in the image quality judgment module, image quality judgment is performed based on one or more features of the image; the features include CDF, HSV and LAB.

[0019] Preferably, the image quality judgment module specifically includes the following submodules: The first submodule of image quality judgment is used to acquire an image and generate CDF features, HSV features and LAB features of the image; The second submodule of image quality judgment is used to fuse the CDF feature, HSV feature and LAB feature to form a first fusion feature; The third submodule of image quality judgment is used to input the first fusion feature into the SVM linear kernel, the SVM polynomial kernel and the SVM Gaussian kernel respectively to form a first classification result, a second classification result and a third classification result respectively; The fourth submodule of image quality judgment is used to process the first classification result, the second classification result and the third classification result through an Adaboost enhanced classifier to form a first image set and a low-illumination image set.

[0020] Preferably, the low-light image enhancement module specifically includes the following submodules: A first low-light image enhancement submodule, configured to convert a first RGB image of the low-light image set into a first color space HSV; A second submodule for low-light image enhancement is configured to smooth the V in the first color space HSV by using a mean filter to form a first brightness component; a low-light image enhancement third submodule, configured to process the first luminance component through guided filtering to form a second luminance component; a fourth submodule for low-light image enhancement, configured to perform logarithmic calculation on the second brightness component to form a reflection component; a fifth submodule for low-light image enhancement, configured to optimize the reflection component through a Gabor filter to form a third brightness component; a sixth low-light image enhancement submodule, configured to synthesize the second luminance component and the third luminance component to form a component V_enhanced; The seventh low-light image enhancement submodule is configured to reorganize the color space according to the component V_enhanced to form a second color space (H, S, V_enhanced), and convert the second color space into a second RGB image (ie, a second image set).

[0021] Preferably, in the model training module, the deep learning model is a Yolo series model.

[0022] Preferably, the deep learning model is a Yolo v11n model; wherein the Yolo v11n model includes a backbone network, a connection layer and an output layer connected in sequence; The backbone network sequentially connects the first convolutional layer, the second convolutional layer, the first C3K2 module, the third convolutional layer, the second C3K2 module, the fourth convolutional layer, the third C3K2 module, the fifth convolutional layer, the fourth C3K2 module, the SPFF module and the C2PSA module; The connection layer comprises a first connection subnetwork and a second connection subnetwork, wherein the first connection subnetwork includes a first UpSample module, a first Concat module, a fifth C3K2 module, a second UpSample module, a second Concat module, and a sixth C3K2 module connected in sequence; and the second connection subnetwork includes a sixth convolutional layer, a third Concat module, a seventh C3K2 module, a seventh convolutional layer, a fourth Concat module, and an eighth C3K2 module connected in sequence; The fifth C3K2 module is connected to the third Concat module, and the sixth C3K2 module is connected to the sixth convolutional layer; The output layer includes a first Detect module, a second Detect module and a third Detect module which do not affect each other; The second C3K2 module is connected to the second Concat module, the third C3K2 module is connected to the first Concat module, the C2PSA module is connected to the first UpSample module and the fourth Concat module, the sixth C3K2 module is connected to the sixth convolutional layer and the first Detect module, the seventh C3K2 module is connected to the second Detect module, and the eighth C3K2 module is connected to the third Detect module.

[0023] More preferably, the Bottleneck module in the C3K2 module is replaced with a lightweight StarBlock module to form an I-C3K2 module.

[0024] Preferably, the first Detect module, the second Detect module and the third Detect module each include two parallel lines, one including two convolutional layers Conv, a two-dimensional convolutional layer Conv2d and a bounding box loss CLoU connected in sequence, and the other including a depth-separable convolutional layer DW-Conv, a convolutional layer Conv, a depth-separable convolutional layer DW-Conv, a convolutional layer Conv, a two-dimensional convolutional layer Conv2d and a classification loss CLSLoss connected in sequence.

[0025] Optionally, one or more of the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the seventh convolutional layer are replaced with the corresponding first depth convolutional layer, the second depth convolutional layer, the third depth convolutional layer, the fourth depth convolutional layer, the fifth depth convolutional layer, the sixth depth convolutional layer and the seventh depth convolutional layer, and / or the first C3K2 module, the second C3K2 module, the third C3K2 module, the first One or more C3K2 modules among the fourth C3K2 module, the fifth C3K2 module, the sixth C3K2 module, the seventh C3K2 module and the eighth C3K2 module are respectively replaced with the corresponding first I-C3K2 module, the second I-C3K2 module, the third I-C3K2 module, the fourth I-C3K2 module, the fifth I-C3K2 module, the sixth I-C3K2 module, the seventh I-C3K2 module and the eighth I-C3K2 module to form an improved Yolo v11 model.

[0026] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an image recognition method as described in any one of the first aspects of the present application.

[0027] In a fourth aspect, the present invention further provides an electronic device comprising: a memory storing a computer program; and a processor communicatively connected to the memory, for executing an image recognition method described in any one of the first aspects of the present application when the computer program is called.

[0028] Compared with the prior art, the present invention has the following obvious outstanding substantial features and significant advantages: The present invention proposes an image recognition method, device, storage medium, and electronic device that address the existing issues of poor low-light image enhancement and low recognition accuracy. The method involves classifying acquired images into a first image set and a low-light image set, and then enhancing the low-light images into a second image set. The first and second image sets are then fed into an improved deep learning model for target detection to train the model. The image to be recognized is then fed into the trained model to generate a recognition result. This significantly improves the enhancement of low-light images, thereby increasing the recognition accuracy of the improved model. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings, which constitute part of the present invention, are provided to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings: Figure 1This is a flow chart of an image recognition method according to a preferred embodiment of the present invention.

[0030] Figure 2 This is a flow chart of an image quality judgment method according to a preferred embodiment of the present invention.

[0031] Figure 3 This is a flow chart of an image enhancement processing method according to a preferred embodiment of the present invention.

[0032] Figure 4 This is a diagram of the Yolo v11 model architecture of a preferred embodiment of the present invention.

[0033] Figure 5 It is a structural diagram of the C3K2 module of a preferred embodiment of the present invention.

[0034] Figure 6 It is a structural diagram of the C2PSA module of a preferred embodiment of the present invention.

[0035] Figure 7 It is a structural diagram of the Detect module of a preferred embodiment of the present invention.

[0036] Figure 8 This is a structural diagram of the I-C3K2 module of a preferred embodiment of the present invention.

[0037] Figure 9 It is a structural diagram of the StarBlock module of the preferred embodiment of the present invention.

[0038] Figure 10 This is a diagram of the improved Yolo v11 model architecture of a preferred embodiment of the present invention.

[0039] Figure 11 It is a structural diagram of an image recognition device according to a preferred embodiment of the present invention.

[0040] Figure 12 It is a structural diagram of an image quality judgment module according to a preferred embodiment of the present invention.

[0041] Figure 13 It is a structural schematic diagram of a low-illumination image enhancement module according to a preferred embodiment of the present invention.

[0042] Reference numerals: 100, image quality judgment module; 200, low illumination image enhancement module; 300, model training module; 400, recognition module; 101, image quality judgment first submodule; 102, image quality judgment second submodule; 103, image quality judgment third submodule; 104, image quality judgment fourth submodule; 201. Low-light image enhancement first submodule; 202. Low-light image enhancement second submodule; 203. Low-light image enhancement third submodule; 204. Low-light image enhancement fourth submodule; 205. Low-light image enhancement fifth submodule; 206. Low-light image enhancement sixth submodule; 207. Low-light image enhancement seventh submodule. DETAILED DESCRIPTION

[0043] The present invention provides an image recognition method, apparatus, storage medium, and electronic device. To clarify the objectives, technical solutions, and effects of the present invention, the present invention is further described below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention.

[0044] It should be noted that the terms "first," "second," and the like in the specification and claims of the present invention and the accompanying drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this manner are interchangeable under appropriate circumstances. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatuses.

[0045] Example 1: like Figures 1-10 As shown, the image recognition method described in this embodiment specifically includes steps S1 to S7.

[0046] Step S1: Acquire images and assess their quality to form a first image set (i.e., a set of images that do not require image enhancement) and a low-light image set. To reduce the impact of low-light images on subsequent recognition, image preprocessing is required. However, over-enhancing images, especially those not in low-light conditions, can introduce additional noise and artifacts, ultimately reducing overall image quality.

[0047] Image quality is assessed using one or more image features (here, image features refer to inherent image features). Specifically, in this embodiment, these features include CDF, HSV, and LAB. Experimental verification has shown that combining CDF, HSV, and LAB features simultaneously for image quality assessment yields high accuracy.

[0048] CDF: The CDF can fully describe the distribution of image pixel brightness values. In low-light images, brightness information is a key factor. By analyzing the CDF, it is possible to accurately determine how many pixels in the image are in darker areas and whether the brightness distribution is uniform. This helps to identify low-quality images where most of the image is covered in darkness or where the brightness distribution is uneven (for example, some areas are too bright and most areas are too dark). For low-light images, the CDF can help determine the span of the image brightness distribution and thus assess whether its contrast is appropriate. If the span of the CDF curve is small, it means that the contrast is low and the image details may be difficult to distinguish. By combining it with the SVM Adaboost model for screening, these images with poor contrast can be eliminated.

[0049] HSV: In low-light environments, the luminance component (V) is an important basis for evaluating image quality. The HSV color space can separate brightness from hue and saturation, making brightness analysis more accurate. By analyzing the luminance component, it is possible to determine whether the overall brightness of the image meets the requirements and whether there are any areas that are too dark or too bright. Even in low-light conditions, the image still contains some color information. Hue (H) and saturation (S) can preserve these color characteristics. This helps to retain images that have low brightness but rich and accurate color information during the screening process, as this color information is very critical in the low-light recognition scenario described in this embodiment.

[0050] LAB: The LAB color space is closer to how the human visual system perceives color. In low-light image quality assessment, this helps more realistically reflect the human eye's perception of image quality. For example, the human eye's perception of brightness is nonlinear, and the brightness component (L) in the LAB color space can better simulate this perceptual characteristic, making the assessment results more consistent with people's subjective feelings. The LAB color space separates brightness and color information. In low-light images, this allows for independent analysis of brightness and color. For the brightness channel, quality assessment can be performed before operations such as brightness enhancement; for the color channel, the accuracy and richness of color can be analyzed, resulting in a more comprehensive assessment of image quality.

[0051] Optionally, the step S1 specifically includes steps S101 to S104.

[0052] Step S101: Acquire an image and generate CDF features, HSV features, and LAB features of the image.

[0053] Step S102: Fusing the CDF feature, HSV feature, and LAB feature to form a first fused feature.

[0054] Step S103: input the first fusion features into the SVM linear kernel, the SVM polynomial kernel and the SVM Gaussian kernel respectively to form a first classification result, a second classification result and a third classification result respectively.

[0055] Step S104: Process the first classification result, the second classification result, and the third classification result through an Adaboost enhanced classifier to form a first image set and a low-illumination image set.

[0056] The above specific implementation steps combine the advantages of support vector machines (SVM) and AdaBoost algorithms, and also have powerful feature extraction and selection capabilities. They can integrate multiple image features and filter out key features. At the same time, they are highly robust to features, which can reduce interference from factors such as noise and improve the accuracy and stability of image evaluation.

[0057] Optionally, before performing image quality assessment, one or at least two of the SVM linear kernel, the SVM polynomial kernel, the SVM Gaussian kernel and the Adaboost enhanced classifier may be trained and optimized to further enhance the accuracy and stability of image assessment.

[0058] Step S2: performing enhancement processing on the low-illumination image set to form a second image set.

[0059] Specifically, the step S2 includes steps S201 to S207.

[0060] Step S201: Convert a first RGB image in the low-illumination image set into a first color space HSV.

[0061] Step S202: Smoothing V in the first color space HSV by using mean filtering to form a first brightness component.

[0062] Step S203: Process the first brightness component through guided filtering to form a second brightness component.

[0063] Step S204: perform logarithmic calculation on the second brightness component to form a reflection component.

[0064] Step S205: Optimize the reflection component using a Gabor filter to form a third luminance component. Specifically, the Gabor filter extracts features at different scales (by adjusting the width of the Gaussian function) and directions (by adjusting the direction parameter θ), capturing multi-scale and multi-directional information in the image.

[0065] Step S206: synthesize the second luminance component and the third luminance component to form a component V_enhanced.

[0066] Step S207 : reorganize the color space according to the component V_enhanced to form a second color space (H, S, V_enhanced), and convert the second color space into a second RGB image (ie, a second image set).

[0067] Step S3: construct a deep learning model, and train the deep learning model using the first image set and the second image set to form a trained deep learning model.

[0068] Specifically, in step S3, the deep learning model is a Yolo series model, such as Yolo v5, Yolo v6, Yolo v8, Yolo v11, improved Yolo v8, etc. Yolo series models. In step S3, the deep learning models that can achieve image target detection are all deep learning models referred to in the present invention. In order to meet the needs of actual engineering applications, and the Yolo series model is relatively mature, this embodiment uses the Yolo series deep learning model as an example for description.

[0069] In this embodiment, Yolo v11n is selected with low computational complexity and fast recognition rate so as to be subsequently deployed on edge devices or mobile terminal devices.

[0070] Among them, the Yolo v11n model includes a backbone network (Baackbone), a connection layer (Neck) and an output layer (Head) connected in sequence; the backbone network is connected in sequence to the first convolutional layer, the second convolutional layer, the first C3K2 module, the third convolutional layer, the second C3K2 module, the fourth convolutional layer, the third C3K2 module, the fifth convolutional layer, the fourth C3K2 module, the SPFF module and the C2PSA module; the connection layer first connection subnetwork and the second connection subnetwork, the first connection subnetwork includes the first UpSample (up sampling) module, the first Concat (splicing) module, the fifth C3K2 module, the second UpSample module, the second Concat module and the sixth C3K2 module connected in sequence; the second connection subnetwork includes the sixth convolutional layer, the third Concat module, the seventh C3K2 module, the fourth C3K2 module, the fifth C3K2 module, the SPFF module and the C2PSA module connected in sequence. Seven convolutional layers, a fourth Concat module and an eighth C3K2 module; the fifth C3K2 module is connected to the third Concat module, and the sixth C3K2 module is connected to the sixth convolutional layer; the output layer includes a first Detect module, a second Detect module and a third Detect module that do not affect each other; the second C3K2 module is connected to the second Concat module, the third C3K2 module is connected to the first Concat module, the C2PSA module is connected to the first UpSample module and the fourth Concat module, the sixth C3K2 module is connected to the sixth convolutional layer and the first Detect module, the seventh C3K2 module is connected to the second Detect module, and the eighth C3K2 module is connected to the third Detect module. The first Detect module, the second Detect module and the third Detect module each include two parallel lines, one including two convolutional layers Conv, a two-dimensional convolutional layer Conv2d and a bounding box loss CLoU connected in sequence, and the other including a depth-wise separable convolutional layer DW-Conv, a convolutional layer Conv, a depth-wise separable convolutional layer DW-Conv, a convolutional layer Conv, a two-dimensional convolutional layer Conv2d and a classification loss CLSLoss connected in sequence.

[0071] Optionally, the Bottleneck module in the C3K2 module is replaced with a lightweight StarBlock module (the StarBlock module structure is as follows Figure 9 As shown, represents star operation), forming an I-C3K2 module.

[0072] Optionally, one or more of the first, second, third, fourth, fifth, sixth, and seventh convolutional layers may be replaced with corresponding depthwise convolutional layers. For example, the first convolutional layer may be replaced with a first depthwise convolutional layer; the second convolutional layer may be replaced with a second depthwise convolutional layer; the third convolutional layer may be replaced with a third depthwise convolutional layer; the fourth convolutional layer may be replaced with a fourth depthwise convolutional layer; the fifth convolutional layer may be replaced with a fifth depthwise convolutional layer; the sixth convolutional layer may be replaced with a sixth depthwise convolutional layer; and the seventh convolutional layer may be replaced with a seventh depthwise convolutional layer.

[0073] Optionally, one or more of the first C3K2 module, the second C3K2 module, the third C3K2 module, the fourth C3K2 module, the fifth C3K2 module, the sixth C3K2 module, the seventh C3K2 module, and the eighth C3K2 module may be replaced with corresponding I-C3K2 modules. For example, the first C3K2 module may be replaced with a first I-C3K2 module; the second C3K2 module may be replaced with a second I-C3K2 module; the third C3K2 module may be replaced with a third I-C3K2 module; the fourth C3K2 module may be replaced with a fourth I-C3K2 module; the fifth C3K2 module may be replaced with a fifth I-C3K2 module; the sixth C3K2 module may be replaced with a sixth I-C3K2 module; the seventh C3K2 module may be replaced with a seventh I-C3K2 module; and the eighth C3K2 module may be replaced with an eighth I-C3K2 module.

[0074] In this embodiment, all convolutional layers and / or modules in the first to seventh convolutional layers and the first to eighth C3K2 modules are replaced to form an improved Yolo v11n model (Yolo v11n-Improve).

[0075] Step S4: Obtain the image to be identified and judge the quality of the image to be identified. If no enhancement is required, input the image to be identified into the trained deep learning model. Otherwise, input the enhanced image to be identified into the trained deep learning model to form a recognition result.

[0076] In actual application scenarios, terminals have limited computing resources. For example, in the case of power inspection robots, the deployment of lightweight deep learning models on mobile terminals is primarily driven by hardware resource limitations, real-time requirements, energy consumption control, multimodal data fusion, and the complexity of the application scenario. These factors include the following:

[0077] 1. Hardware resource limitations Power inspection robots typically use embedded processors or specialized chips, which have limited computing power and storage space. For example, some robots may only be equipped with low-end ARM processors, which have computing power far lower than high-performance servers. Lightweight models can better adapt to these hardware conditions by reducing the number of parameters and computational complexity.

[0078] 2. Real-time requirements Power inspection tasks require high real-time performance, requiring robots to quickly process sensor data and make decisions. For example, when detecting power equipment faults, lightweight models can significantly reduce inference time, thus meeting real-time requirements.

[0079] 3. Energy consumption control Power inspection robots typically rely on batteries, making energy consumption control crucial. The low computational complexity of lightweight models can reduce chip heat generation and energy consumption, thereby extending battery life. For example, using lightweight models can significantly reduce the hardware resources required for model forward computation, thereby reducing energy consumption.

[0080] 4. Multimodal Data Fusion Power inspection scenarios are complex, making accurate representation difficult with single-modal data. Lightweight models can better support multimodal data fusion, increasing the information content and discriminability of feature representation. For example, in all-weather substation inspections, feature-level fusion of visible light and infrared images allows lightweight models to maintain high detection accuracy even in adverse weather conditions.

[0081] 5. Complexity of application scenarios Power inspection robots need to operate in complex and changing environments, such as outdoors and at high altitudes. Lightweight models can be more flexibly deployed on different devices and adapt to complex application scenarios.

[0082] In summary, the image recognition method described in this embodiment can be deployed on a variety of manned or unmanned mobile terminal devices, such as power inspection robots, drones, unmanned vehicles, and unmanned boats. It achieves efficient, fast, and low-energy operation with limited hardware resources, while meeting the needs of complex application scenarios. This embodiment targets image recognition in low-light scenarios.

[0083] In addition, experiments were conducted based on the image recognition method described in this embodiment. Common metrics used in the field of this technology for evaluating model quality were selected, including P, R, mAP50, mAP50-95, Params, GFLOPs, and FPS.

[0084] P (Precision) represents the proportion of samples predicted by the model as positive that are actually positive, and is used to measure the accuracy of the model's positive predictions. A higher P value indicates a more reliable prediction of positive samples. The formula is: P = TP / (TP + FP); where TP (True Positive) is the number of true positives, or samples correctly predicted as positive by the model; FP (False Positive) is the number of false positives, or samples incorrectly predicted as positive by the model.

[0085] R (Recall) represents the proportion of samples that are actually positive that are correctly predicted as positive by the model. It measures the model's coverage of positive samples. A higher R value indicates that the model is able to detect more positive samples. The formula is: R = TP / (TP + FN); where FN (False Negative) is the number of false negatives, that is, the number of samples that the model incorrectly predicted as negative.

[0086] mAP50 (mean Average Precision at IoU=0.5) represents the average precision (AP) calculated with an IoU (Intersection over Union) threshold of 0.5. AP is the area under the precision-recall curve. mAP is the average AP across multiple categories or IoU thresholds. It measures the average precision of a model at different recall rates and is a commonly used comprehensive performance metric in object detection.

[0087] mAP50-95 (mean Average Precision at IoU = 0.5:0.95) represents the average precision calculated over an IoU threshold range of 0.5 to 0.95 (in steps of 0.05). It is stricter than mAP50 and can more comprehensively evaluate model performance under different precision requirements. It is used to measure the overall performance of a model under high-precision requirements.

[0088] Params (Parameters) represents the number of trainable parameters in the model, usually expressed in millions (106), and is used to measure the complexity and storage requirements of the model. The larger the number of parameters, the more expressive the model may be, but the computational cost and storage requirements will also be higher.

[0089] GFLOPs (Giga Floating-Point Operations per Second) represents the number of floating-point operations required for each forward pass of the model, expressed in billions, and is used to measure the computational complexity of the model. The lower the GFLOPs, the more efficient the model is, making it suitable for running on resource-constrained devices.

[0090] FPS (Frames Per Second) indicates the number of image frames a model can process per second. It is used to measure the real-time performance of a model and its running speed in practical applications. The higher the FPS, the better the real-time performance of the model, making it suitable for scenarios such as real-time video processing.

[0091] This experiment was completed using the NVIDIA GeForce RTX 3090 GPU, and the experimental results are shown in Table 1.

[0092] Table 1 Comparison table of image recognition comparison experiment results

[0093] As shown in Table 1, for the precision P, the Yolov11n model achieved a precision of 99.4%, and the Yolov11n-Improve model achieved a precision of 99.8%. The Yolov11n-Improve model has a higher precision, which means that among all samples predicted as positive, the Yolov11n-Improve model has a larger proportion of correct predictions, reducing the number of false positives.

[0094] For recall R, the Yolov11n model achieved a recall of 100%, while the Yolov11n-Improve model achieved a recall of 99.8%. Although Yolov11n's recall was slightly higher, the Yolov11n-Improve model's recall was also very high, only 0.2% lower than the Yolov11n model. This indicates that the Yolov11n-Improve model is almost as effective as the Yolov11n model in identifying all positive samples.

[0095] For the average precision mAP50, the mAP50 of the Yolov11n model reached 99.5%, and the mAP50 of the Yolov11n-Improve model reached 99.5%. On the mAP50 indicator, the two models performed the same, indicating that when the IoU threshold is 0.5, the average precision of Yolov11n-Improve is comparable to that of Yolov11n.

[0096] For the average precision mAP50-95, the Yolov11n model achieved an mAP50-95 of 98.8%, and the Yolov11n-Improve model achieved an mAP50-95 of 98.2%. Although the Yolov11n model was slightly better in mAP50-95, the performance of the Yolov11n-Improve model was quite close, only 0.6% lower. This shows that under a stricter IoU threshold range, the performance of Yolov11n-Improve is still very good.

[0097] For the parameter amount (Params), the parameter amount of the Yolov11n model is 2.58×10 6 , the number of parameters of the Yolov11n-Improve model is only 1.74×10 6 , the number of parameters of Yolov11n-Improve is significantly reduced, which means that the model is more lightweight, occupies less storage space, and is easy to deploy on resource-constrained devices.

[0098] In terms of computational complexity (GFLOPs), the computational complexity of the Yolov11n model is 6.3, and the computational complexity of the Yolov11n-Improve model is 4.2. The computational complexity of the Yolov11n-Improve model is lower, indicating that it has lower computational complexity and requires fewer computing resources, thereby improving the operation efficiency of the model.

[0099] In terms of frames per second (FPS), the FPS of the Yolov11n model reached 1148.5, and the FPS of the Yolov11n-Improve model reached 1398.1. The FPS of the Yolov11n-Improve model was significantly improved, which means that the model has faster processing speed and is more suitable for real-time application scenarios.

[0100] In summary, the Yolov11n-Improve model demonstrates significant advantages in the following areas: higher precision and fewer false positives; significantly fewer parameters, resulting in a more lightweight model; lower computational complexity, improving operational efficiency; and significantly faster processing, making it more suitable for real-time applications. Although the Yolov11n model slightly outperforms the other two metrics in terms of recall (R) and mean average precision (mAP50-95), the Yolov11n-Improve model performs very closely and exhibits significant advantages in other key metrics. These advantages make the Yolov11n-Improve model a well-balanced model between performance and efficiency, making it particularly suitable for resource-constrained applications or those requiring fast processing.

[0101] Example 2: like Figure 11-13 As shown, the image recognition device described in this embodiment specifically includes an image quality judgment module 100, a low-illumination image enhancement module 200, a model training module 300 and a recognition module 400.

[0102] Image quality assessment module 100 is configured to acquire images, assess their quality, and form a first image set (i.e., a set of images that do not require image enhancement) and a low-light image set. Specifically, image quality assessment is performed based on one or more image features, including CDF, HSV, and LAB.

[0103] Specifically, the image quality judgment module includes a first image quality judgment submodule 101 , a second image quality judgment submodule 102 , a third image quality judgment submodule 103 and a fourth image quality judgment submodule 104 .

[0104] The first image quality judgment submodule 101 is used to acquire an image and generate CDF features, HSV features and LAB features of the image.

[0105] The second image quality judgment submodule 102 is configured to fuse the CDF feature, the HSV feature, and the LAB feature to form a first fused feature.

[0106] The third image quality judgment submodule 103 is used to input the first fusion features into the SVM linear kernel, the SVM polynomial kernel and the SVM Gaussian kernel respectively to form a first classification result, a second classification result and a third classification result respectively.

[0107] The fourth image quality judgment submodule 104 is configured to process the first classification result, the second classification result, and the third classification result through an Adaboost enhanced classifier to form a first image set and a low-illumination image set.

[0108] The low-illumination image enhancement module 200 is configured to perform enhancement processing on the low-illumination image set to form a second image set.

[0109] Specifically, the training image enhancement module includes a low-light image enhancement first submodule 201, a low-light image enhancement second submodule 202, a low-light image enhancement third submodule 203, a low-light image enhancement fourth submodule 204, a low-light image enhancement fifth submodule 205, a low-light image enhancement sixth submodule 206 and a low-light image enhancement seventh submodule 207.

[0110] The low-illumination image enhancement first submodule 201 is configured to convert the first RGB image of the low-illumination image set into a first color space HSV.

[0111] The second low-illumination image enhancement submodule 202 is configured to smooth the V in the first color space HSV by using a mean filter to form a first brightness component.

[0112] The third low-illumination image enhancement submodule 203 is configured to process the first brightness component through guided filtering to form a second brightness component.

[0113] The fourth low-illumination image enhancement submodule 204 is configured to perform logarithmic calculation on the second brightness component to form a reflection component.

[0114] The fifth low-illumination image enhancement submodule 205 is configured to optimize the reflection component through a Gabor filter to form a third brightness component.

[0115] The sixth low-illumination image enhancement submodule 206 is configured to synthesize the second luminance component and the third luminance component to form a component V_enhanced.

[0116] The seventh low-illumination image enhancement submodule 207 is configured to reorganize the color space according to the component V_enhanced to form a second color space (H, S, V_enhanced), and convert the second color space into a second RGB image, ie, a second image set.

[0117] The model training module 300 is used to construct a deep learning model and train the deep learning model using the first image set and the second image set to form a trained deep learning model.

[0118] Specifically, in the model training module 300, the deep learning model is a Yolo series model, such as Yolo v5, Yolo v6, Yolo v8, Yolo v11, improved Yolo v8, and the like.

[0119] In this embodiment, the deep learning model is a Yolo v11n model; wherein the Yolo v11n model includes a backbone network, a connection layer and an output layer connected in sequence; the backbone network is connected in sequence to a first convolutional layer, a second convolutional layer, a first C3K2 module, a third convolutional layer, a second C3K2 module, a fourth convolutional layer, a third C3K2 module, a fifth convolutional layer, a fourth C3K2 module, an SPFF module and a C2PSA module; the connection layer includes a first connection subnetwork and a second connection subnetwork, the first connection subnetwork includes a first UpSample module, a first Concat module, a fifth C3K2 module, a second UpSample module, a second Concat module and a sixth C3K2 module connected in sequence; the second connection subnetwork includes a sixth convolutional layer, a third Concat module, a seventh C3K2 module, a seventh convolutional layer, a fourth Concat module and a block and the eighth C3K2 module; the fifth C3K2 module is connected to the third Concat module, and the sixth C3K2 module is connected to the sixth convolutional layer; the output layer includes a first Detect module, a second Detect module and a third Detect module that do not affect each other; the second C3K2 module is connected to the second Concat module, the third C3K2 module is connected to the first Concat module, the C2PSA module is connected to the first UpSample module and the fourth Concat module, the sixth C3K2 module is connected to the sixth convolutional layer and the first Detect module, the seventh C3K2 module is connected to the second Detect module, and the eighth C3K2 module is connected to the third Detect module. The first Detect module, the second Detect module and the third Detect module each include two parallel lines, one including two convolutional layers Conv, a two-dimensional convolutional layer Conv2d and a bounding box loss CLoU connected in sequence, and the other including a depth-wise separable convolutional layer DW-Conv, a convolutional layer Conv, a depth-wise separable convolutional layer DW-Conv, a convolutional layer Conv, a two-dimensional convolutional layer Conv2d and a classification loss CLSLoss connected in sequence.

[0120] Optionally, the Bottleneck module in the C3K2 module is replaced with a lightweight StarBlock module to form an I-C3K2 module.

[0121] Optionally, one or more of the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the seventh convolutional layer are replaced with the corresponding first depth convolutional layer, the second depth convolutional layer, the third depth convolutional layer, the fourth depth convolutional layer, the fifth depth convolutional layer, the sixth depth convolutional layer and the seventh depth convolutional layer, and / or the first C3K2 module, the second C3K2 module, the third C3K2 module, the first One or more C3K2 modules among the fourth C3K2 module, the fifth C3K2 module, the sixth C3K2 module, the seventh C3K2 module and the eighth C3K2 module are respectively replaced with the corresponding first I-C3K2 module, the second I-C3K2 module, the third I-C3K2 module, the fourth I-C3K2 module, the fifth I-C3K2 module, the sixth I-C3K2 module, the seventh I-C3K2 module and the eighth I-C3K2 module to form an improved Yolo v11 model.

[0122] The recognition module 400 is used to obtain the image to be recognized and judge the quality of the image to be recognized. If no enhancement is required, the image to be recognized is input into the trained deep learning model. Otherwise, the enhanced image to be recognized is input into the trained deep learning model to form a recognition result.

[0123] While the specific embodiments of the present invention have been described in detail above, these are merely exemplary and the present invention is not limited thereto. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, any equivalent changes and modifications made without departing from the spirit and scope of the present invention are intended to be encompassed within the scope of the present invention.

Claims

1. An image recognition method, characterized in that: The specific steps include: Acquiring images, judging the quality of the images, and forming a first image set and a low-illumination image set; performing enhancement processing on the low-illumination image set to form a second image set; Constructing a deep learning model, and training the deep learning model using the first image set and the second image set to form a trained deep learning model; Obtain the image to be identified and judge the quality of the image to be identified. If no enhancement is required, input the image to be identified into the trained deep learning model. Otherwise, input the enhanced image to be identified into the trained deep learning model to form a recognition result.

2. An image recognition method according to claim 1, characterized in that: Image quality is judged by one or more features of the image; wherein the features include CDF, HSV and LAB.

3. The image recognition method according to claim 1, wherein: Acquiring images, judging the quality of the images, and forming a first image set and a low-illumination image set specifically includes the following steps: Acquire an image and generate CDF features, HSV features, and LAB features of the image; Performing feature fusion on the CDF feature, the HSV feature, and the LAB feature to form a first fused feature; Inputting the first fusion features into the SVM linear kernel, the SVM polynomial kernel and the SVM Gaussian kernel respectively to form a first classification result, a second classification result and a third classification result respectively; The first classification result, the second classification result, and the third classification result are processed by an Adaboost enhanced classifier to form a first image set and a low-illumination image set.

4. The image recognition method according to claim 1, wherein: Performing enhancement processing on the low-illumination image set to form a second image set specifically includes the following steps: Converting the first RGB image of the low-light image set into a first color space HSV; Smoothing V in the first color space HSV by using a mean filter to form a first brightness component; Processing the first luminance component through guided filtering to form a second luminance component; performing logarithmic calculation on the second brightness component to form a reflection component; Optimizing the reflection component through a Gabor filter to form a third brightness component; Combining the second luminance component and the third luminance component to form a component V_enhanced; The color space is reorganized according to the component V_enhanced to form a second color space H, S, V_enhanced, and the second color space is converted into a second RGB image.

5. The image recognition method according to claim 1, wherein: The deep learning model is a Yolov11n model; wherein the Yolov11n model includes a backbone network, a connection layer and an output layer connected in sequence; The backbone network sequentially connects the first convolutional layer, the second convolutional layer, the first C3K2 module, the third convolutional layer, the second C3K2 module, the fourth convolutional layer, the third C3K2 module, the fifth convolutional layer, the fourth C3K2 module, the SPFF module and the C2PSA module; The connection layer comprises a first connection subnetwork and a second connection subnetwork, wherein the first connection subnetwork includes a first UpSample module, a first Concat module, a fifth C3K2 module, a second UpSample module, a second Concat module, and a sixth C3K2 module connected in sequence; and the second connection subnetwork includes a sixth convolutional layer, a third Concat module, a seventh C3K2 module, a seventh convolutional layer, a fourth Concat module, and an eighth C3K2 module connected in sequence; The fifth C3K2 module is connected to the third Concat module, and the sixth C3K2 module is connected to the sixth convolutional layer; The output layer includes a first Detect module, a second Detect module and a third Detect module which do not affect each other; The second C3K2 module is connected to the second Concat module, the third C3K2 module is connected to the first Concat module, the C2PSA module is connected to the first UpSample module and the fourth Concat module, the sixth C3K2 module is connected to the sixth convolutional layer and the first Detect module, the seventh C3K2 module is connected to the second Detect module, and the eighth C3K2 module is connected to the third Detect module.

6. An image recognition method according to claim 5, characterized in that: The Bottleneck module in the C3K2 module is replaced with a lightweight StarBlock module to form an I-C3K2 module.

7. An image recognition method according to claim 5, characterized in that: The first Detect module, the second Detect module and the third Detect module each include two parallel lines, one including two convolutional layers Conv, a two-dimensional convolutional layer Conv2d and a bounding box loss CLoU connected in sequence, and the other including a depth-wise separable convolutional layer DW-Conv, a convolutional layer Conv, a depth-wise separable convolutional layer DW-Conv, a convolutional layer Conv, a two-dimensional convolutional layer Conv2d and a classification loss CLSLoss connected in sequence.

8. An image recognition device, characterized in that: Specifically, it includes the following modules: An image quality judgment module, configured to acquire images, judge the quality of the images, and form a first image set and a low-illumination image set to be enhanced; a low-illumination image enhancement module, configured to perform enhancement processing on the low-illumination image set to form a second image set; A model training module, configured to construct a deep learning model and train the deep learning model using the first image set and the second image set to form a trained deep learning model; The recognition module is used to obtain the image to be recognized and judge the quality of the image to be recognized. If no enhancement is required, the image to be recognized is input into the trained deep learning model. Otherwise, the enhanced image to be recognized is input into the trained deep learning model to form a recognition result.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the image recognition method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements an image recognition method according to any one of claims 1 to 7 when executing the computer program.

Images