Image recognition method and device, computer readable storage medium and processor

Abstract

This invention discloses an image recognition method, apparatus, computer-readable storage medium, and processor. The method includes: acquiring an image to be recognized; inputting the image to be recognized into a trained YOLO-Final model, whereby the YOLO-Final model identifies the target in the image. The YOLO-Final model includes at least an original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model. The final layer neural network module indicates the number of neurons and weight parameters for identifying the target. This invention solves the technical problem of missed detection of certain targets by existing YOLO series target detection neural network models.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to an image recognition method, apparatus, computer-readable storage medium, and processor. Background Technology

[0002] In recent years, with breakthroughs in deep learning technology, performance in target recognition and detection from camera images has been greatly improved. Therefore, deep learning technology has been increasingly applied to target detection in autonomous driving camera perception systems. Classic deep learning models for target detection include Faster R-CNN and the YOLO series. The YOLO series, due to its higher real-time performance, is more widely used in real-time detection systems like autonomous driving systems. From YOLO-v3 to YOLO-v4 / YOLO-v5 in 2020, newer versions have been developed.

[0003] Through in-depth practical research on the application of YOLO series deep learning neural network models in target perception of autonomous driving camera images, several shortcomings were discovered:

[0004] 1. The neural network training of the YOLO series models uses a large number of neurons and weight parameters for target classification, thus weakening the training and detection of target presence / absence. In autonomous driving applications, target detection is the most important, while classification errors are acceptable.

[0005] 2. The YOLO series models employ a multi-scale feature map detection method to accurately detect bounding boxes of targets of different sizes during neural network training. Typically, feature maps of three image sizes are used for target detection. Therefore, during training, a large number of neurons and weight parameters are distributed across bounding boxes of different sizes, thus reducing the weight of neurons that assess the presence of the target. However, in autonomous driving applications, the importance of target presence (i.e., detection) far outweighs the need to minimize bounding box errors.

[0006] There is currently no effective solution to the above problems. Summary of the Invention

[0007] This invention provides an image recognition method, apparatus, computer-readable storage medium, and processor to at least solve the technical problem of missed detection of certain targets by existing YOLO series target detection neural network models.

[0008] According to one aspect of the present invention, an image recognition method is provided, comprising: acquiring an image to be recognized; inputting the image to be recognized into a trained YOLO-Final model, wherein the YOLO-Final model identifies a target in the image to be recognized, wherein the YOLO-Final model includes at least an original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model, the final layer neural network module being used to indicate the number of neurons and weight parameters of the target to be recognized.

[0009] Optionally, inputting the image to be identified into a trained YOLO-Final model, and having the YOLO-Final model identify the target in the image to be identified, includes: inputting the image to be identified into the original YOLO neural network model to obtain feature maps at different scales; fusing the feature maps at different scales to obtain a fused feature map; inputting the fused feature map into the final layer neural network module to obtain a label feature map; determining multiple candidate points in the label feature map, wherein the candidate points are the locations of the center points of potential targets; inputting the multiple candidate points and the candidate regions extracted from the output bounding boxes / classification features of the original YOLO neural network model into the post-processing module of the original YOLO neural network model for target confirmation; and outputting the confirmed target.

[0010] Optionally, before acquiring the image to be identified, the method further includes: constructing a yolo-final model based on the original yolo neural network model; and training the yolo-final model.

[0011] Optionally, constructing a YOLO-final model based on the original YOLO neural network model includes: determining the original YOLO neural network model; adding a final layer neural network module at the end of the original YOLO neural network model to generate the YOLO-final model, wherein the number of convolutional neural network layers in the final layer neural network module includes at least one of the following: 4 layers or 7 layers.

[0012] Optionally, training the YOLO-Final model includes: determining the center coordinates of the known target bounding boxes labeled on the training image; obtaining a label feature map based on a preset feature map value and a preset radius of the center coordinates; inputting the training image into the YOLO-Final model, and having the YOLO-Final model determine the training result; determining the training loss based on the label feature map and the training result; and iteratively optimizing the network by taking the back derivative of the training loss using a conventional deep learning loss function until the label feature map and the training result match.

[0013] According to another aspect of the present invention, an image recognition device is also provided, comprising: an acquisition unit for acquiring an image to be recognized; and a recognition unit for inputting the image to be recognized into a trained YOLO-Final model, wherein the YOLO-Final model recognizes the target in the image to be recognized, wherein the YOLO-Final model includes at least an original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model, and the final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be recognized.

[0014] Optionally, the recognition unit includes: a first processing subunit, used to input the image to be recognized into the original YOLO neural network model to obtain feature maps at different scales; a fusion subunit, used to fuse the feature maps at different scales to obtain a fused feature map; a second processing subunit, used to input the fused feature map into the final layer neural network module to obtain a label feature map; a first determining subunit, used to determine multiple candidate points of the label feature map, wherein the candidate points are the locations of the center points of potential targets; a confirmation subunit, used to input the multiple candidate points and the candidate regions extracted from the output bounding boxes / classification features of the original YOLO neural network model into the post-processing module of the original YOLO neural network model for target confirmation; and an output subunit, used to output the confirmed target.

[0015] Optionally, the device further includes: a construction unit for constructing a yolo-final model based on the original yolo neural network model; and a training unit for training the yolo-final model.

[0016] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the image recognition method described in any one of the above.

[0017] According to another aspect of the present invention, a processor is also provided, the processor being configured to run a program, wherein the program, when running, executes the image recognition method described in any one of the foregoing embodiments.

[0018] In this embodiment of the invention, an image to be identified is acquired; the image to be identified is input into a trained YOLO-Final model, and the YOLO-Final model identifies the target in the image to be identified. The YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model. The final layer neural network module is used to indicate the number of neurons and weight parameters for identifying the target. By identifying the target in the image through the trained YOLO-Final model, the goal of quickly and accurately identifying the target in the image to be identified is achieved, thereby improving the target detection rate and solving the technical problem of missed detection of certain targets by existing YOLO series target detection neural network models. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0020] Figure 1 This is a flowchart of an image recognition method according to an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of an image recognition device according to an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] Example 1

[0025] According to an embodiment of the present invention, an embodiment of an image recognition method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0026] Figure 1 This is a flowchart of an image recognition method according to an embodiment of the present invention, such as... Figure 1 As shown, the image recognition method includes the following steps:

[0027] Step S102: Obtain the image to be recognized;

[0028] Step S104: Input the image to be recognized into the trained YOLO-Final model, and the YOLO-Final model recognizes the target in the image to be recognized. The YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added at the end of the original YOLO neural network model. The final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be recognized.

[0029] As an optional embodiment, by adding a final layer neural network module at the end of the original YOLO neural network model, the proportion of neurons and weight parameters for detecting the presence or absence of targets by cameras in the entire model can be increased, thereby achieving the effect of improving the target detection rate.

[0030] As an optional embodiment, the added final layer neural network module can highlight the center position of the target region and gradually reduce the weights from the center position outwards, thereby forming a feature map in which the target center region on the image reaches a maximum value, thus better detecting the target and reducing the missed detection of some targets by the original YOLO network.

[0031] It should be noted that the number of neurons and weight parameters mentioned above can be set according to the needs of the application scenario. The addition of the final layer neural network module results in the overall network having more neurons and weight parameters, which increases the proportion of target detection in the entire neural network model.

[0032] Through the above steps, an image to be identified can be acquired and then input into a trained YOLO-Final model. The YOLO-Final model can then identify the target in the image. The YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model. The final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be identified. By using the trained YOLO-Final model to identify the target in the image, the goal of quickly and accurately identifying the target in the image is achieved, thereby improving the target detection rate and solving the technical problem of missed detection of certain targets by existing YOLO series target detection neural network models.

[0033] Optionally, the image to be recognized is input into a trained YOLO-Final model, and the YOLO-Final model identifies the target in the image by: inputting the image to be recognized into the original YOLO neural network model to obtain feature maps at different scales; fusing the feature maps at different scales to obtain a fused feature map; inputting the fused feature map into the final layer neural network module to obtain a label feature map; determining multiple candidate points in the label feature map, wherein the candidate points are the locations of the center points of potential targets; inputting the multiple candidate points and the output bounding boxes / candidate regions extracted from the classification features of the original YOLO neural network model into the post-processing module of the original YOLO neural network model for target confirmation; and outputting the confirmed target.

[0034] As an optional implementation, firstly, the image to be identified is input into the original YOLO neural network model and processed through multiple layers to obtain feature maps at different scales. Secondly, the feature maps at different scales are fused and input into the final layer neural network module to obtain a label feature map. Then, multiple local maxima points are obtained from this label feature map, which are the locations of the center points of potential targets. Finally, the candidate points and the candidate regions extracted from the output bounding boxes / classification features of the original YOLO neural network model are fed together into the post-processing module (the conventional algorithm module of the original YOLO framework) of the original YOLO neural network model for final target selection and confirmation, and then the confirmed targets are output. Through the above implementation method, the accuracy of target recognition or detection can be effectively improved.

[0035] Optionally, before acquiring the image to be recognized, the above method further includes: constructing a yolo-final model based on the original yolo neural network model; and training the yolo-final model.

[0036] As an optional implementation, the YOLO-Final model is built on the basis of the original YOLO neural network model, and the built YOLO-Final model is trained to obtain the optimal parameters, thereby improving the recognition accuracy of the YOLO-Final model.

[0037] Optionally, constructing a YOLO-final model based on the original YOLO neural network model includes: determining the original YOLO neural network model; adding a final layer neural network module at the end of the original YOLO neural network model to generate the YOLO-final model, wherein the number of convolutional neural network layers in the final layer neural network module includes at least one of the following: 4 layers or 7 layers.

[0038] As an optional embodiment, the above-described YOLO-final model is obtained by adding a final layer neural network module to the end of the original YOLO neural network model. In specific implementations, the number of convolutional neural network layers in the above-described final layer neural network module includes, but is not limited to, 4 layers, 7 layers, etc., and can also be other numbers of layers, which will not be elaborated here.

[0039] Optionally, training the YOLO-Final model includes: determining the center coordinates of the known target bounding boxes based on the known target bounding boxes labeled in the training image; obtaining the label feature map based on the preset feature map values ​​and preset radius of the center coordinates; inputting the training image into the YOLO-Final model, and having the YOLO-Final model determine the training result; determining the training loss based on the label feature map and the training result; and iteratively optimizing the network by taking the back derivative of the training loss using a conventional deep learning loss function until the label feature map and the training result match.

[0040] As an optional implementation, during the training of the YOLO-Final model, the center coordinates of the known target bounding boxes labeled on the training images need to be determined. Next, based on the preset feature map values ​​and preset radius of the center coordinates, a label feature map is obtained. The training images are then input into the YOLO-Final model, which determines the training result. Further, based on the label feature map and the training result, the training loss is determined. Finally, the training loss is iteratively optimized using a conventional deep learning loss function until the label feature map and the training result match. Through the above implementation, a YOLO-Final model that can accurately identify image targets can be obtained. It should be noted that the above training process optimizes the relevant parameters of the YOLO-Final model.

[0041] The following is a detailed description of an optional embodiment of the present invention.

[0042] In an optional embodiment of the present invention, the basic neural network model still follows the YOLO-V5 framework. On the basis of the YOLO-V5 framework, a final layer neural network module is added to train weights for the existence of the target in the image and its center position.

[0043] The optional training process is as follows:

[0044] First, calculate the center coordinates of the known target bounding boxes labeled in the image used for training;

[0045] Secondly, the center coordinate feature map value is set to 1, and it radiates outwards with a certain radius R. R = minimum(box_Width, box_Height) / 2, that is, R is the minimum value of the width and height of the target bounding box, divided by 2, and its feature map value decays from 1 at the center to 0 at a distance R from the center, thus forming a map with the maximum value at the center of the target and a certain slope in the central area. This label feature map is named label-object-feature.

[0046] The final layer neural network module is composed of multiple convolutional operations, for example, it is composed of 7 convolutional layers. After the feature map of the image to be detected is extracted by the original YOLO network, it is then subjected to multiple convolutional operations of the final layer neural network module. Finally, the loss (i.e. error) is calculated with the label-object-feature map.

[0047] Then, the network is iteratively optimized by taking the inverse derivative of the loss function commonly used in deep learning until the image to be detected passes through the entire neural network and the output from the final layer best matches the label-object-feature. At this point, the training of the overall network is terminated.

[0048] The optional prediction process is as follows:

[0049] After training, the neurons in each layer of the final layer neural network module have obtained optimal parameters. When the image to be recognized is input from the input end of the YOLO-Final network:

[0050] First, after the calculations of each layer of the original YOLO-V5 model before the final module, feature maps of different scales are obtained;

[0051] Secondly, feature maps of different scales are fused and input into the final layer neural network module for neural network operations, outputting a label-object-feature;

[0052] Then, from the label-object-feature feature map, obtain multiple local maximum points, which are the locations of the center points of the potential targets;

[0053] Finally, the candidate points and the candidate regions extracted from the output bounding boxes / classification features of the original YOLO-V5 framework are fed into the post-processing module of the original YOLO-V5 framework for final target selection and confirmation.

[0054] Output the confirmed target.

[0055] As an optional implementation, to achieve better neural network training results, the final layer neural network module uses a relatively large number of convolutional neural network layers (7 layers); for the feature values ​​of the target center region, the decay is from the center point 1 to 0 at the radius R, where R is taken as a large value. However, this processing will increase the number of neural network layers, the number of neurons, and the weight parameters, thereby increasing the computational load and the requirements for computer storage space.

[0056] As another optional embodiment, a design with reduced complexity is adopted, which reduces the computational load by about half, but the performance is only slightly reduced. The specific implementation method and the overall process are exactly the same as the first embodiment, with the following two differences: only 4 convolutional neural network layers are used as the final layer neural network module; when calculating the feature value of the target center region, the attenuation radius is reduced: R = minimum(box_Width, box_Height) / 3, that is: R is taken as the minimum value of the width and height of the target bounding box, and then divided by 3 or 4, etc., so that the R value is smaller, thereby reducing the area involved in the loss calculation.

[0057] In the above embodiments of the present invention, by improving the added final layer neural network module, the proportion of target detection in the entire neural network model is increased. This allows more neurons and weight parameters in the overall network to contribute to target detection, slightly weakening the accuracy of target classification and target bounding boxes. Thus, with a small sacrifice in classification and bounding box accuracy, the detection of target presence and its center location can be significantly improved, which is highly compatible with the perception requirements of autonomous driving cameras. As for the sacrificed bounding box accuracy, it can be compensated for by data fusion correction using 3D point clouds obtained from other sensors in the autonomous driving perception system, such as LiDAR, thereby compensating for the loss of bounding box accuracy in camera image detection.

[0058] It should be noted that in the specific implementation, 5100 real-time road traffic images were used, including real street scenes, vehicles, pedestrians, bicycles, motorcycles, and tricycles. For the original YOLO system, we used the latest YOLO-v5 neural network model released in 2020 and compared the results with the YOLO-final model of this invention.

[0059] Target detection rate = Total number of detected targets / Total number of all labeled targets;

[0060] The target detection rate of the YOLO-v5 model reached 86.6%; while the target detection rate of the YOLO-final model of this invention reached 95.3%.

[0061] Therefore, by employing the final module of this invention, the target detection rate for autonomous driving is significantly improved from 86.6% to 95.3%. It should be noted that some smaller or heavily occluded targets were not detected by the YOLO-v5 model, but were successfully detected by the YOLO-final model of this invention.

[0062] Example 2

[0063] According to another aspect of the present invention, an image recognition device is also provided. Figure 2This is a schematic diagram of an image recognition device according to an embodiment of the present invention, such as... Figure 2 As shown, the image recognition device includes an acquisition unit 22 and a recognition unit 24. The image recognition device will now be described in detail.

[0064] Acquisition unit 22 is used to acquire the image to be recognized; recognition unit 24 is connected to the acquisition unit 22 and is used to input the image to be recognized into the trained YOLO-Final model, so that the YOLO-Final model can recognize the target in the image to be recognized. The YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added at the end of the original YOLO neural network model. The final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be recognized.

[0065] In the above embodiments of the present invention, the image recognition device can identify the target in the image to be recognized by the trained YOLO-Final model, thereby achieving the purpose of quickly and accurately identifying the target in the image to be recognized, thus realizing the technical effect of improving the target detection rate, and solving the technical problem of the YOLO series target detection neural network model in the prior art missing detection of some targets.

[0066] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, they can be implemented in the following ways: the above modules can be located in the same processor; and / or, the above modules can be located in different processors in any combination.

[0067] It should be noted that the acquisition unit 22 and the identification unit 24 mentioned above correspond to steps S102 to S104 in Embodiment 1. The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above units, as part of the device, can be executed in a computer system such as a set of computer-executable instructions.

[0068] Optionally, the aforementioned recognition unit includes: a first processing subunit for inputting the image to be recognized into the original YOLO neural network model to obtain feature maps at different scales; a fusion subunit for fusing the feature maps at different scales to obtain a fused feature map; a second processing subunit for inputting the fused feature map into the final layer neural network module to obtain a label feature map; a first determination subunit for determining multiple candidate points in the label feature map, wherein the candidate points are the locations of the center points of potential targets; a confirmation subunit for inputting the multiple candidate points and the candidate regions extracted from the output bounding boxes / classification features of the original YOLO neural network model into the post-processing module of the original YOLO neural network model for target confirmation; and an output subunit for outputting the confirmed target.

[0069] Optionally, the above apparatus further includes: a construction unit for constructing a yolo-final model based on the original yolo neural network model; and a training unit for training the yolo-final model.

[0070] Optionally, the above-mentioned building unit includes: a second determining subunit for determining the original YOLO neural network model; and a generating subunit for adding a final layer neural network module at the end of the original YOLO neural network model to generate a YOLO-final model, wherein the number of convolutional neural network layers in the final layer neural network module includes at least one of the following: 4 layers or 7 layers.

[0071] Optionally, the training unit includes: a third determination subunit, used to determine the center coordinates of the known target bounding boxes based on the known target bounding boxes labeled in the training image; a third processing subunit, used to obtain the label feature map based on the preset feature map value and preset radius of the center coordinates; a fourth determination subunit, used to input the training image into the YOLO-Final model, and the YOLO-Final model determines the training result; a fifth determination subunit, used to determine the training loss based on the label feature map and the training result; and a fourth processing subunit, used to perform back-derivative of the training loss using a conventional loss function in deep learning, iteratively optimizing the network until the label feature map and the training result match.

[0072] Example 3

[0073] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform any of the above-described image recognition methods.

[0074] Optionally, in this embodiment, the computer-readable storage medium may be located in any computer terminal in a group of computer terminals in a computer network, and / or in any mobile terminal in a group of mobile terminals, and the computer-readable storage medium includes a stored program.

[0075] Optionally, during program execution, the device containing the computer-readable storage medium is controlled to perform the following functions: acquire an image to be recognized; input the image to be recognized into a trained YOLO-Final model, and have the YOLO-Final model recognize the target in the image to be recognized, wherein the YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added at the end of the original YOLO neural network model, and the final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be recognized.

[0076] Example 4

[0077] According to another aspect of the present invention, a processor is also provided for running a program, wherein the program executes the image recognition method described above.

[0078] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps: acquiring an image to be recognized; inputting the image to be recognized into a trained YOLO-Final model, and having the YOLO-Final model recognize the target in the image to be recognized. The YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model. The final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be recognized.

[0079] The present invention also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program having the following method steps: acquiring an image to be recognized; inputting the image to be recognized into a trained YOLO-Final model, and having the YOLO-Final model recognize the target in the image to be recognized, wherein the YOLO-Final model includes at least an original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model, and the final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be recognized.

[0080] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0081] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0082] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0083] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0085] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0086] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An image recognition method, characterized in that, include: Acquire the image to be recognized; The image to be identified is input into a trained YOLO-Final model, which identifies the target in the image. The YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model. The final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be identified. The image to be identified is input into a trained YOLO-Final model, and the YOLO-Final model identifies the target of the image to be identified, including: The image to be identified is input into the original YOLO neural network model to obtain feature maps at different scales; The feature maps of different scales are fused to obtain a fused feature map; The fused feature map is input into the final layer neural network module to obtain the label feature map; Determine multiple candidate points in the label feature map, wherein the candidate points are the locations of the center points of potential targets; The multiple candidate points and the candidate regions extracted from the output bounding boxes / classification features of the original YOLO neural network model are input into the post-processing module of the original YOLO neural network model for target confirmation. The confirmed target will be output.

2. The method according to claim 1, characterized in that, Before acquiring the image to be identified, the method further includes: The YOLO-Final model is constructed based on the original YOLO neural network model. Train the YOLO-Final model.

3. The method according to claim 2, characterized in that, Building the YOLO-Final model based on the original YOLO neural network model includes: Determine the original YOLO neural network model; A final layer neural network module is added to the end of the original YOLO neural network model to generate the YOLO-final model, wherein the number of convolutional neural network layers of the final layer neural network module includes at least one of the following: 4 layers or 7 layers.

4. The method according to claim 2, characterized in that, Training the YOLO-Final model includes: Based on the known target bounding boxes labeled in the training images, determine the center coordinates of the known target bounding boxes; Based on the preset feature map value and preset radius of the center coordinates, a label feature map is obtained; The training images are input into the YOLO-Final model, and the YOLO-Final model determines the training results. Based on the label feature map and the training results, determine the training loss; The training loss is inversely derived using a conventional deep learning loss function, and the network is iteratively optimized until the label feature map matches the training result.

5. An image recognition device, characterized in that, include: The acquisition unit is used to acquire the image to be recognized; The recognition unit is used to input the image to be recognized into a trained YOLO-Final model, and the YOLO-Final model recognizes the target in the image to be recognized. The YOLO-Final model includes at least the original YOLO neural network model and a final layer neural network module added to the end of the original YOLO neural network model. The final layer neural network module is used to indicate the number of neurons and weight parameters of the target to be recognized. The identification unit includes: The first processing subunit is used to input the image to be identified into the original YOLO neural network model to obtain feature maps at different scales; The fusion subunit is used to fuse the feature maps of different scales to obtain a fused feature map; The second processing subunit is used to input the fused feature map into the final layer neural network module to obtain the label feature map; The first determining subunit is used to determine multiple candidate points of the label feature map, wherein the candidate point is the location of the center point of the potential target; The confirmation subunit is used to input the multiple candidate points and the candidate regions extracted from the output bounding box / classification features of the original YOLO neural network model into the post-processing module of the original YOLO neural network model for target confirmation. The output subunit is used to output the confirmed target.

6. The apparatus according to claim 5, characterized in that, The device further includes: Building units are used to construct YOLO-Final models based on the original YOLO neural network models. The training unit is used to train the YOLO-Final model.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the image recognition method according to any one of claims 1 to 4.

8. A processor, characterized in that, The processor is used to run a program, wherein the program executes the image recognition method according to any one of claims 1 to 4.

Images