Image recognition method, electronic device, storage medium, program product and vehicle
By adding a first classification model to the perception model, optimizing and iterating the training set to process target recognition, the problem of pedestrian misidentification was solved, stable and reliable target recognition was achieved, vehicle safety and user experience were improved, and costs were reduced.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-31
- Publication Date
- 2026-03-31
AI Technical Summary
In existing technologies, pedestrian recognition misidentification processing is costly and not stable or reliable enough, which leads to a reduction in vehicle driving safety and user experience.
A first classification model is added to the perceptual model. The target classification result is determined by the target detection box and the first classification model. The first classification model is optimized to correct incorrect target classification results. The model is then optimized by iteratively training the training set to achieve stable and reliable target recognition.
It improves the accuracy of suppressing false recognition, reduces software and hardware costs, and enhances vehicle driving safety and user experience.
Smart Images
Figure CN121767947A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to an image recognition method, electronic device, computer-readable storage medium, computer program product, and vehicle. Background Technology
[0002] Image recognition is crucial for vehicle driving safety. Taking pedestrian recognition as an example, the handling of false pedestrian recognition usually involves combining multimodal recognition results to suppress false pedestrian detections. This means that the post-processing simultaneously receives the recognition results from the LiDAR and the recognition results from the forward-looking camera and performs fusion processing to reduce emergency braking caused by false pedestrian detections from the forward-looking camera.
[0003] However, the above methods require high hardware and software recognition costs, are not stable and reliable enough in handling misidentification, and have low accuracy in suppressing misidentification, which reduces vehicle driving safety and user experience. Summary of the Invention
[0004] The present invention aims to solve at least one of the technical problems existing in the prior art.
[0005] Therefore, one objective of this invention is to propose an image recognition method that reduces the hardware and software costs of image recognition, improves the accuracy of suppressing false recognition, and enhances the stability and reliability of image recognition, thereby improving vehicle driving safety and user experience.
[0006] Therefore, a second objective of the present invention is to provide an electronic device.
[0007] Therefore, a third objective of the present invention is to provide a computer-readable storage medium.
[0008] Therefore, a fourth objective of this invention is to provide a computer program product.
[0009] Therefore, the fifth objective of this invention is to provide a vehicle.
[0010] To achieve the above objectives, an embodiment of the first aspect of the present invention provides an image recognition method, the method comprising: determining a target classification result based on a target detection box and a first classification model; optimizing the first classification model based on the target classification result to obtain an optimized first classification model; and outputting a target object based on the optimized first classification model and the target detection box.
[0011] According to the image recognition method of this invention, a first classification model is added to the perception model to suppress false recognition. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of falsely identified targets, improve the accuracy of suppressing false recognition, thereby improving vehicle driving safety and user experience. Moreover, it does not require LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0012] In some embodiments, optimizing the first classification model based on the target classification result includes: obtaining false positive classification results in the target classification result; performing pre-refresh processing on the false positive classification results to obtain pre-refresh results; determining the false positive classification dataset based on the pre-refresh results; and optimizing the first classification model based on the false positive classification dataset.
[0013] In some embodiments, determining the false alarm classification dataset based on the pre-brush result includes: when the pre-brush result is correct, setting the false alarm classification result as a non-target object; when the pre-brush result is incorrect, labeling the false alarm classification result to obtain labeled classification objects; and determining the false alarm classification dataset based on the non-target objects and the labeled classification objects.
[0014] In some embodiments, before determining the target classification result based on the target detection box and the first classification model, the method includes: acquiring image data of the mobile device; and acquiring the target detection box based on the image data.
[0015] In some embodiments, obtaining the target detection box based on the image data includes: determining the tensor form of the image data; determining the detection box based on the tensor form and the perceptual model; and obtaining the target detection box based on the detection box.
[0016] In some embodiments, determining the target classification result based on the target detection box and the first classification model includes: scaling the target detection box to obtain a scaled target detection box; and determining the target classification result based on the scaled target detection box and the first classification model.
[0017] To achieve the above objectives, a second aspect of the present invention provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores an image recognition program executable by the at least one processor, and when the image recognition program is executed by the at least one processor, the at least one processor performs the image recognition method as described in the above embodiments.
[0018] According to the electronic device of the present invention, the image recognition method of the above embodiments is used to suppress misidentification by adding a first classification model on the basis of the perception model. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of misidentified targets, improve the accuracy of suppressing misidentification, thereby improving vehicle driving safety and user experience. Moreover, it does not require LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0019] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing an image recognition program, which, when executed by a processor, causes a device equipped with the image recognition program to implement the image recognition method as described in the above embodiments.
[0020] To achieve the above objectives, a fourth aspect of the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the image recognition method as described in the above embodiments.
[0021] According to the computer program product of the present invention, a first classification model is added to the perception model to suppress false recognition. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of falsely identified targets, improve the accuracy of suppressing false recognition, thereby improving vehicle driving safety and user experience. Moreover, it does not require LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0022] To achieve the above objectives, a fifth aspect of the present invention provides a vehicle comprising: electronic equipment as described in the above embodiments.
[0023] According to an embodiment of the present invention, the vehicle is equipped with the electronic device described in the above embodiment. By adding a first classification model to the perception model, false recognition is suppressed. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is placed into a classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects. This achieves stable and reliable handling of falsely identified targets, improves the accuracy of suppressing false recognition, thereby improving vehicle driving safety and user experience. Furthermore, it eliminates the need for LiDAR sensors and complex post-processing rule code, reducing hardware and software costs.
[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0025] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of an image recognition method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an image recognition method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a perception model processing according to an embodiment of the present invention; Figure 4 This is a flowchart of an image recognition method according to another embodiment of the present invention; Figure 5 This is a block diagram of an electronic device according to an embodiment of the present invention; Figure 6 This is a vehicle block diagram according to an embodiment of the present invention.
[0026] Figure label: 90. Forward-looking 2D image; 91. Dynamic perception model; 92. Optimized first classification model; Processor 101; Memory 102; Electronic device 100; Vehicle 99. Detailed Implementation
[0027] The embodiments described with reference to the accompanying drawings are exemplary, and the embodiments of the present invention are described in detail below.
[0028] Image recognition is crucial for vehicle driving safety. Taking pedestrian recognition as an example, pedestrians are vulnerable road users and are key targets for identification and avoidance in driver assistance systems. However, due to the varied postures and rich features of pedestrians, high demands are placed on the pedestrian recognition model of the forward-facing camera. False detection of pedestrians in driver assistance systems can easily trigger AEB (Autonomous Emergency Braking), causing safety hazards and a poor driving experience.
[0029] In related technologies, the processing of pedestrian misidentification usually involves combining multimodal recognition results to suppress false pedestrian detections. This means that the post-processing simultaneously receives the recognition results from the LiDAR and the recognition results from the forward-looking camera and performs fusion processing to reduce emergency braking and other issues caused by false pedestrian detections from the forward-looking camera.
[0030] However, the aforementioned methods require high hardware and software recognition costs, necessitating complex post-processing code logic to apply various rules to the recognition results from different modal sensors. Pedestrian false alarm suppression based on multi-sensor fusion recognition requires inputting the LiDAR recognition model results into the fusion module, and performing association matching and state estimation analysis of 2D and 3D bounding boxes to determine whether the target is a pedestrian. Furthermore, the handling of false alarms is not stable or reliable enough, resulting in low accuracy in suppressing false alarms, which reduces vehicle driving safety and the user experience.
[0031] Therefore, the image recognition method of this invention suppresses misidentification by adding a first classification model to the perception model. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of misidentified targets, improve the accuracy of suppressing misidentification, thereby improving vehicle driving safety and user experience. Moreover, it eliminates the need for LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0032] The following is combined Figures 1-4 The image recognition method according to an embodiment of the present invention is described.
[0033] like Figure 1 The diagram shown is a flowchart of an image recognition method according to an embodiment of the present invention. The image recognition method of this embodiment includes at least steps S1-S3.
[0034] Step S1: Determine the target classification result based on the target detection box and the first classification model.
[0035] In this embodiment, for the target detection box, the forward-looking camera acquisition module uses forward-looking wide-angle and narrow-angle cameras to acquire image data of the vehicle's surroundings in real time during the vehicle's driving process. In the image preprocessing module, the video stream read by the camera is decoded and converted into a tensor form that the perception model can read and recognize. The tensor form is input into the perception model, which can detect various dynamic and static object elements in the vehicle driving scene, infer and output the labeled box of the detected object, i.e., the detection box. Assuming the detection target is a pedestrian, the target detection box, i.e., the pedestrian detection box, can be obtained from the labeled box. For the first classification model, which is an iterative model trained on a false alarm classification dataset, it will further judge and classify the pedestrian detection box of the perception model, filtering out some common false alarms, i.e., misidentification. Taking a pedestrian as an example, the pedestrian detection box is input into the first classification model. The first classification model will determine whether the pedestrian detection box is a real pedestrian and obtain the target classification result. For example, the target is not a pedestrian, or the target is a pedestrian. The result is output to the downstream so that the downstream can provide feedback based on the target classification result, thereby optimizing the first classification model.
[0036] Step S2: Optimize the first classification model based on the target classification result to obtain the optimized first classification model.
[0037] In this embodiment, the correctness of the target classification result is judged manually. For target classification results that are judged incorrectly, the data is sent for labeling and pre-labeling with a large model. The incorrectly labeled target classification results are automatically corrected. The results of the pre-labeling with the large model are judged. For target classification results that fail to be automatically corrected, manual correction labeling is performed. The automatically corrected target classification results and the manually corrected labels are put into the training dataset of the first classification model. The first classification model is optimized. Through continuous iteration, a stable and accurate optimized first classification model is obtained.
[0038] Step S3: Output the target object based on the optimized first classification model and the target detection box.
[0039] In this embodiment, a stable and accurate optimized first classification model, obtained through continuous iteration, is used. The target detection box is input into the optimized first classification model to improve the reliability of pure visual forward perception or recognition by adding an optimized first classification model to the recognition result of the perception model. The perception result is further classified and suppressed, and misidentified targets are handled stably and reliably, improving the accuracy of suppressing misidentification. As the training dataset continues to increase, the classification ability of the model gradually increases, ultimately providing more stable and reliable target recognition results for complex target scenes, thereby improving vehicle driving safety and user experience. Moreover, it eliminates the need for LiDAR sensors and complex post-processing rule code, reducing hardware and software costs. According to the image recognition method of this invention, a first classification model is added to the perception model to suppress false recognition. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of falsely identified targets, improve the accuracy of suppressing false recognition, thereby improving vehicle driving safety and user experience. Moreover, it does not require LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0040] In some embodiments, the method of optimizing the first classification model based on the target classification result includes: obtaining the false positive classification result in the target classification result; performing a pre-brushing process on the false positive classification result to obtain a pre-brushing result; determining the false positive classification dataset based on the pre-brushing result; and optimizing the first classification model based on the false positive classification dataset.
[0041] In this embodiment, false positive classification results are obtained from the target classification results. For example, during the development process, after obtaining the target classification results, the downstream will provide feedback on the incorrect judgments output by the first classification model. A manual review is then conducted to determine whether the target classification result is correct, i.e., whether the target is not a pedestrian, or whether the target is a pedestrian and the result is incorrectly identified. False positive classification results are then obtained. Correctly identified results from the manually determined target classification results are added to the classification model's training dataset to optimize the data in the training dataset, making the model's training results more accurate. Pre-brushing processing, such as large model pre-brushing annotation, involves re-identifying and annotating the false positive classification results, thus pre-brushing the false positive classification results. For example, initially, while the model performance is still in the improvement stage, data that is not correctly classified needs to be further accumulated through large-scale model pre-processing and manual annotation. Specifically, the false positive classification result data is sent for annotation, large-scale model pre-processing is performed to obtain pre-processing results, the correctness of the pre-processing results is judged and processed accordingly to obtain the false positive classification dataset. The annotation results of the targets in the false positive classification dataset are all correct, so as to achieve further judgment and processing of the false positive classification results. Based on the data in the false positive classification dataset where the annotation results are all correct, the first classification model is optimized to obtain an accurate and reliable optimized first classification model.
[0042] In some embodiments, determining the false alarm classification dataset based on the pre-brush result includes: when the pre-brush result is correct, setting the false alarm classification result as a non-target object; when the pre-brush result is incorrect, labeling the false alarm classification result to obtain labeled classification objects; and determining the false alarm classification dataset based on the non-target objects and labeled classification objects.
[0043] In this embodiment, the pre-brush result includes incorrect and correct annotations. An incorrect annotation might be that the image data indicates the target is not a pedestrian, but the pre-brush annotation indicates the target is a pedestrian. A correct annotation might be that the image data indicates the target is not a pedestrian, and the pre-brush annotation also indicates the target is not a pedestrian. When the pre-brush result is a correct annotation, the false positive classification result is set to a non-target object to automatically correct the false positive classification result. When the pre-brush result is an incorrect annotation, the false positive classification result is annotated, for example, manually correcting the incorrect annotation to obtain an annotated classification object, thus correcting the false positive classification result where automatic correction failed. The non-target object and the annotated classification object are used as a false positive classification dataset to incorporate the corrected data into the training data of the first classification model for training, allowing the first classification model to relearn and iterate to improve the model's correct classification ability.
[0044] In some embodiments, before determining the target classification result based on the target detection box and the first classification model, the method includes: acquiring image data of the mobile device; and acquiring the target detection box based on the image data.
[0045] In this embodiment, the mobile device is a device that needs to perform target recognition, such as a vehicle or a robot. Taking a vehicle as an example, the forward-facing camera acquisition module uses forward-facing wide-angle and narrow-angle cameras to acquire image data around the vehicle in real time. When acquiring data in real time, the state of the vehicle is not limited. As long as the vehicle has a target recognition requirement, real-time acquisition can be performed. For example, the vehicle's state can be driving, parking, or starting, etc., to obtain the data source for image recognition. The target detection box marks the target to be identified in the image in the form of a box. The acquired image data is processed to obtain the target detection box, so as to realize the preliminary target recognition of the acquired image and output it.
[0046] In some embodiments, obtaining a target detection box based on image data includes: determining the tensor form of the image data; determining a detection box based on the tensor form and a perceptual model; and obtaining the target detection box based on the detection box.
[0047] In an embodiment, such as Figure 2 The diagram illustrates an image recognition method according to an embodiment of the present invention. The perception model is a multi-task visual perception model capable of recognizing and labeling detected objects in the form of bounding boxes. In the image preprocessing module, the video stream read by the camera is decoded and converted into a tensor form that the perception model can read and recognize, such as a forward-looking 2D (two-dimensional) image 90. The forward-looking 2D image 90 is a tensor form, preparing the perception model to recognize and label detected objects. Figure 3 The diagram illustrates a perception model processing method according to an embodiment of the present invention. Image data in tensor form is input into the perception model (e.g., dynamic perception model 91). The perception model contains an AI (Artificial Intelligence) neural network that can detect various dynamic and static object elements in a vehicle driving scene, infer and output the bounding boxes of the detected objects, i.e., detection boxes. For example, it outputs pedestrian detection boxes, vehicle detection boxes, traffic light detection boxes, and other detection boxes to prepare for obtaining target detection boxes. Target detection boxes are obtained based on the detection boxes. For example, if the focus of identification during vehicle movement is pedestrians, then the target is the pedestrian, and the target detection box is the pedestrian detection box. Obtaining the pedestrian detection boxes within the detection boxes prepares for determining the target classification result.
[0048] In some embodiments, determining the target classification result based on the target detection box and the first classification model includes: scaling the target detection box to obtain a scaled target detection box; and determining the target classification result based on the scaled target detection box and the first classification model.
[0049] In this embodiment, taking a pedestrian detection box as an example, the pedestrian detection box is a 2D detection box. It is scaled by 0.9-1.1 times to obtain a scaled pedestrian detection box. During the data loading stage of training the first classification model, there is also a corresponding scaling to adapt to the dataset of the first classification model. The scaled region is used as input to the first classification model. The first classification model is an iterative model trained on a false positive classification dataset. It further classifies the pedestrian detection boxes from the perception model, filtering out some common false positives (false identifications). The scaled pedestrian detection box is input into the first classification model, which determines whether it represents a real pedestrian. The resulting target classification result is, for example, "the target is not a pedestrian" or "the target is a pedestrian." This result is output to the downstream system so that the downstream system can provide feedback based on the target classification result, thereby optimizing the first classification model.
[0050] The following is for reference. Figure 4 The vehicle control method of this invention will be described in detail below.
[0051] like Figure 4 The diagram shown is a flowchart of an image recognition method according to another embodiment of the present invention. The method of this embodiment includes at least steps S10-S24.
[0052] Step S10: Acquire image data of the vehicle while it is driving on the road.
[0053] Step S11: Determine the tensor form of the image data.
[0054] Step S12: Determine the detection box based on the tensor form and the perception model.
[0055] Step S13: Obtain the target detection box based on the detection box.
[0056] Step S14: Scale the target detection box to obtain the scaled target detection box.
[0057] Step S15: Obtain the first classification model.
[0058] Step S16: Determine the target classification result based on the scaled target detection box and the first classification model.
[0059] Step S17: Obtain the false positive classification results from the target classification results.
[0060] Step S18: Perform pre-refresh processing on the false alarm classification results to obtain the pre-refresh results.
[0061] Step S19: Determine if the pre-flash result is correct. If yes, proceed to step S21; otherwise, proceed to step S20.
[0062] Step S20: Label the false positive classification results to obtain labeled classified objects.
[0063] Step S21: Set the false alarm classification result as a non-target object.
[0064] Step S22: Determine the false alarm classification dataset based on non-target objects and labeled classification objects.
[0065] Step S23 yields the optimized first classification model.
[0066] Step S24: Output the target object based on the optimized first classification model and the target detection box.
[0067] According to the image recognition method of this invention, a first classification model is added to the perception model to suppress false recognition. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of falsely identified targets, improve the accuracy of suppressing false recognition, thereby improving vehicle driving safety and user experience. Moreover, it does not require LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0068] The following is for reference. Figure 5 An electronic device that describes an embodiment of the present invention.
[0069] like Figure 5 The diagram shown is a block diagram of an electronic device according to an embodiment of the present invention. An electronic device 100 according to an embodiment of the present invention includes: at least one processor 101; and a memory 102 communicatively connected to the at least one processor 101; wherein the memory 102 stores an image recognition program executable by the at least one processor 101, and when the image recognition program is executed by the at least one processor 101, the at least one processor 101 performs the image recognition method as described in the above embodiment.
[0070] In some embodiments, the processor 101 is configured to: optimize a first classification model based on a target classification result, including: obtaining false positive classification results in the target classification result; performing pre-brushing processing on the false positive classification results to obtain pre-brushing results; determining a false positive classification dataset based on the pre-brushing results; and optimizing the first classification model based on the false positive classification dataset.
[0071] In some embodiments, the processor 101 is configured to: determine a false alarm classification dataset based on the pre-brush result, including: when the pre-brush result is correct, setting the false alarm classification result as a non-target object; when the pre-brush result is incorrect, labeling the false alarm classification result to obtain labeled classification objects; and determining the false alarm classification dataset based on the non-target objects and the labeled classification objects.
[0072] In some embodiments, the processor 101 is configured to: before determining the target classification result based on the target detection box and the first classification model, include: acquiring image data of the vehicle during road travel; and acquiring the target detection box based on the image data.
[0073] In some embodiments, the processor 101 is configured to: obtain a target detection box based on image data, including: determining a tensor form of the image data; determining a detection box based on the tensor form and a perceptual model; and obtaining the target detection box based on the detection box.
[0074] In some embodiments, the processor 101 is configured to: determine a target classification result based on the target detection box and a first classification model, including: scaling the target detection box to obtain a scaled target detection box; and determining the target classification result based on the scaled target detection box and the first classification model.
[0075] According to the electronic device 100 of the present invention, the image recognition method of the above embodiment is used to suppress misidentification by adding a first classification model on the basis of the perception model. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of misidentified targets, improve the accuracy of suppressing misidentification, thereby improving vehicle driving safety and user experience. Moreover, it does not require LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0076] The following describes a computer-readable storage medium according to embodiments of the present invention.
[0077] The computer-readable storage medium of the present invention stores an image recognition program, and when the image recognition program is executed by a processor, the device on which the image recognition program is installed implements the image recognition method as described in the above embodiments.
[0078] The following describes a computer program product based on an embodiment of the present invention.
[0079] The computer program product of this invention includes a computer program, which, when executed by a processor, implements the image recognition method as described in the above embodiments.
[0080] According to the computer program product of the present invention, a first classification model is added to the perception model to suppress false recognition. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is put into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects, so as to achieve stable and reliable handling of falsely identified targets, improve the accuracy of suppressing false recognition, thereby improving vehicle driving safety and user experience. Moreover, it does not require LiDAR sensors and complex post-processing rule code, reducing software and hardware costs.
[0081] like Figure 6 The diagram shown is a vehicle block diagram according to an embodiment of the present invention. The vehicle 99 includes an electronic device 100 as described in the above embodiment.
[0082] According to an embodiment of the present invention, a vehicle 99 is equipped with the electronic device 100 of the above embodiment. By adding a first classification model to the perception model, false recognition is suppressed. The target detection box output by the perception model is input into the first classification model to obtain the target classification result. The correctness of the target classification result is judged, and incorrect target classification results are corrected. The corrected target classification result is placed into the classification training set. The first classification model is trained and optimized using the continuously iterative classification training set to obtain an accurate and reliable optimized first classification model. The target detection box is input into the optimized first classification model to obtain accurately identified target objects. This achieves stable and reliable handling of falsely identified targets, improves the accuracy of suppressing false recognition, thereby improving vehicle driving safety and user experience. Furthermore, it eliminates the need for LiDAR sensors and complex post-processing rule code, reducing hardware and software costs.
[0083] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.
[0084] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.
Claims
1. An image recognition method, characterized in that, include: The target classification result is determined based on the target detection box and the first classification model; The first classification model is optimized based on the target classification result to obtain the optimized first classification model; The target object is output based on the optimized first classification model and the target detection box.
2. The image recognition method according to claim 1, characterized in that, Optimizing the first classification model based on the target classification result includes: Obtain the false positive classification results from the target classification results; The false alarm classification results are pre-refreshed to obtain pre-refresh results; The false alarm classification dataset is determined based on the pre-brush results; The first classification model is optimized based on the false positive classification dataset.
3. The image recognition method according to claim 2, characterized in that, The step of determining the false positive classification dataset based on the pre-brush results includes: When the pre-brush result is correct, the false alarm classification result is set as a non-target object; When the pre-brush result is incorrect, the false classification result is marked to obtain the labeled classified object; The false alarm classification dataset is determined based on the non-target objects and the labeled classification objects.
4. The image recognition method according to claim 1, characterized in that, Before determining the target classification result based on the target detection box and the first classification model, the method further includes: Acquire image data from mobile devices; The target detection box is obtained based on the image data.
5. The image recognition method according to claim 4, characterized in that, The step of obtaining the target detection box based on the image data includes: Determine the tensor form of the image data; The detection box is determined based on the tensor form and the perception model; The target detection box is obtained based on the detection box.
6. The image recognition method according to claim 1, characterized in that, The step of determining the target classification result based on the target detection box and the first classification model includes: The target detection box is scaled to obtain a scaled target detection box; The target classification result is determined based on the scaled target detection box and the first classification model.
7. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores an image recognition program that can be executed by the at least one processor. When the image recognition program is executed by the at least one processor, the at least one processor performs the image recognition method as described in any one of claims 1-6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an image recognition program, which, when executed by a processor, causes a device equipped with the image recognition program to implement the image recognition method as described in any one of claims 1-6.
9. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the image recognition method as described in any one of claims 1-6.
10. A vehicle comprising: The electronic device as claimed in claim 7.