Method for detecting metal surface defects of automotive upholstery based on self-adaptive enhancement algorithm

Abstract

The invention provides a method for detecting metal surface defects of automotive upholstery based on an adaptive enhancement algorithm. The method comprises the following steps: firstly, preprocessing and manually marking existing metal product surface image data and making a data set; then randomly extracting a plurality of samples from the data set to train a first base model; performing self-inspection on the data set by using the first base model, positioning samples with missing detection and misjudgment, increasing the possibility that the misdetection samples are extracted in the next extraction, and iterating the process to obtain a plurality of base models; and finally, in an actual defect detection process, integrating detection results of the base models to obtain a final result. According to the method, a self-adaptive enhancement algorithm is combined, and a strategy of integrating a plurality of base models is adopted, so that missing detection and misjudgment samples in a detection process can be dynamically corrected, a more stable and efficient detection result is provided, and a high recall rate and a low misjudgment rate can be kept in the application field of metal surface defect detection of industrial automotive upholstery.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial metal surface defect detection, and in particular to a method for detecting metal surface defects of automobile interior decoration parts based on an adaptive enhancement algorithm. Background Art

[0002] From the proposal of the Fourth Industrial Revolution to the practice of manufacturing transformation in countries around the world, intelligent manufacturing has broken the limitations of traditional production methods by integrating new-generation information technologies such as the Internet of Things, big data, cloud computing and artificial intelligence, and established a highly automated, interconnected and intelligent production system. It has become the core driving force for industrial upgrading. In the metal manufacturing industry, the development of intelligent manufacturing marks a major shift in the traditional manufacturing model towards high efficiency, flexibility and precision. Relying on intelligent equipment, advanced processes and real-time data analysis, companies can significantly improve production efficiency, optimize resource utilization, and meet customized and small-batch production needs. In addition, intelligent manufacturing has also promoted the digitalization of the production process and runs through the entire product life cycle, including design, manufacturing, logistics and after-sales service, forming a complete intelligent industrial chain; especially in automobile manufacturing, the production cost of interior parts is usually high, and once defects occur, rework and scrap will bring additional time and economic costs. If defects are not discovered in time, a large number of unqualified products may flow into the market, which will require rework or recycling. This will not only waste resources but also may affect the production cycle. Through intelligent defect detection technology, detection efficiency can be improved, human omissions can be reduced, and it can be ensured that every interior part on the production line meets quality standards, thereby reducing scrap rate and rework costs.

[0003] In this context, metal surface defect detection, as an important part of quality control, is directly related to product reliability and market competitiveness. Traditional surface inspection methods usually rely on manual visual inspection or contact instruments, which are not only inefficient but also easily interfered with by human factors, resulting in frequent missed inspections or misjudgments. With the acceleration of production pace and the increase in product complexity, this inspection method has been unable to meet the high standards of modern manufacturing. Therefore, the application of intelligent inspection technology has become the key to solving this problem.

[0004] Currently, thanks to the deep empowerment of artificial intelligence (AI), metal surface defect detection has undergone a comprehensive upgrade from inefficient to highly efficient, from static to dynamic, and from manual to intelligent. During the metal surface defect detection process, non-contact inspection technology based on high-definition industrial cameras, laser scanning, and deep learning algorithms enables accurate surface data collection and analysis of large quantities of metal products in a short period of time. AI not only rapidly identifies surface defects such as scratches, cracks, and pits, but also intelligently classifies and locates complex texture defects, supporting real-time detection and dynamic adjustment. Through big data analysis, the system optimizes detection models for different product types and process requirements, enabling continuous improvement. These technological advantages significantly improve the efficiency and accuracy of defect detection, significantly reducing scrap rates and rework costs. Furthermore, intelligent inspection technology can be deeply integrated with production lines, enabling online inspection and automated sorting, promptly removing substandard products and ensuring high quality standards for every product leaving the factory. More importantly, the traceability of defect detection data throughout the entire process provides data support for enterprise quality management systems, helping to identify process bottlenecks and continuously improve production processes. This lays a solid foundation for metal manufacturers to gain both technological and quality advantages in the fiercely competitive market.

[0005] While artificial intelligence algorithms (especially deep learning) have demonstrated impressive performance in surface defect detection, their adaptability to diverse metal types, surface treatments, and defect types remains limited. Training requires extensive labeled data, which can be challenging to collect and high-quality training data is often difficult to obtain in certain areas. Furthermore, the algorithms' recognition rate for complex and subtle defects still needs improvement, especially in the presence of background noise or reflected light, which can reduce detection accuracy. Therefore, a new approach is needed that can adapt to changing inspection scenarios while ensuring high accuracy and reliability. Summary of the Invention

[0006] To address the challenges of existing technologies, this paper designs a method for metal surface defect detection on automotive interior parts based on an adaptive enhancement algorithm. First, existing metal product surface image data is preprocessed and manually annotated to create a dataset. Next, a first base model is trained by randomly selecting several samples from the dataset. This first base model then performs a self-check on the dataset to locate missed and misclassified samples, increasing the likelihood that these misclassified samples will be detected in the next sampling iteration. This process is repeated to generate multiple base models. Finally, during actual defect detection, the detection results from these base models are integrated to obtain the final result. Compared to training a single model, this method maintains a high recall rate and a low false positive rate in the application of metal surface defect detection on industrial automotive interior parts.

[0007] The technical solution of the present invention is:

[0008] A method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm, characterized by comprising the following steps:

[0009] Step 1: Acquire images;

[0010] Acquire the surface image of a real metal workpiece similar to the metal product to be tested;

[0011] Step 2: Data cleaning;

[0012] Step 3: Manual labeling;

[0013] Determine the type of defects to be detected and manually annotate the pre-processed real metal workpiece surface images to generate a real metal workpiece surface image dataset;

[0014] Step 4: Train the model;

[0015] Step 4.1: Assign the same initial weight to each sample in the real metal workpiece surface image dataset;

[0016] Step 4.2: Randomly extract 80% of the total number of real metal workpiece surface image datasets according to weights as the training dataset for the first base model;

[0017] Step 4.3: Train the base model;

[0018] Step 4.4: Defect detection;

[0019] Use the trained base model to detect real metal workpiece surface image datasets, and compare the detection results with the expert manual annotation results to filter out incorrect detection samples;

[0020] Step 4.5: Increase the weight of the falsely detected samples, and keep the weights of the remaining samples unchanged;

[0021] Step 4.6: Repeat steps 4.2 to 4.5 to iteratively train multiple base models. The number of base models depends on the specific situation.

[0022] Step 5: Performance testing;

[0023] The defect detection performance of each base model is tested individually, and the base models that meet the performance standards are screened out, while the base models with poor performance are discarded; the detection results of the remaining base models are integrated, and the final result is obtained through an integrated voting mechanism.

[0024] Furthermore, the step 2 is to pre-process the real metal workpiece surface image obtained in step 1, specifically including:

[0025] First, the part of the image that does not contain the metal workpiece surface is deleted, and then the remaining part is cut into a size suitable for defect detection. Then, the grayscale space transformation of the metal workpiece surface image is performed to avoid the overall image being too dark or the grayscale value distribution being too concentrated.

[0026] Furthermore, the real metal workpiece surface image dataset generated in step 3 includes the type and location information of the defects and the real metal workpiece surface image.

[0027] Furthermore, in step 4.2, when extracting the training data set, the number of defective samples and non-defective samples is ensured to be balanced. The detailed process is as follows:

[0028] During extraction, a number is randomly generated from 0 to 1 each time. Each sample corresponds to a range. The range length of a sample is the result of normalizing the sample weight. The ranges of all samples are connected end to end without overlapping, covering all ranges between 0 and 1. When the randomly generated number is within the range corresponding to a sample, the sample is placed in the training data set. The number of samples in the final training data set is equal to the number of times the above extraction is repeated.

[0029] Furthermore, in step 4.3, the base model structure is a convolutional neural network, which consists of an input, a backbone network, a neck, and a detection head;

[0030] The input part converts the original image into a data format suitable for base model processing, and performs necessary data enhancement and preprocessing operations;

[0031] The backbone network is responsible for extracting basic visual features from the input image and capturing low-level features and high-level semantic features in the image through multi-layer convolution operations.

[0032] The neck part further processes and integrates the features generated by the backbone network to generate multi-scale feature maps to meet the requirements of target detection for different target sizes;

[0033] The detection head part predicts the location and category of the target based on the feature map generated by the neck, and outputs the category probability of each target and the coordinate information of the bounding box.

[0034] Furthermore, in step 4.4, the error detection samples include missed detection samples and misjudged samples;

[0035] The missed detection samples are samples that were marked as defects by experts during manual annotation, but were not detected by the model;

[0036] The misjudged samples are samples in which defects are not marked when the experts manually annotate them, but the model detects defects in the samples.

[0037] Furthermore, in step 4.5, when the weight reaches an upper threshold, it will no longer increase, so as to avoid the serious imbalance of the extracted training data set due to excessive weight of some samples after multiple false detections.

[0038] Furthermore, the indicators for evaluating the defect detection performance of the base model in step 5 mainly include the recall rate Recall and the false positive rate FPR, as follows:

[0039]

[0040] Where,

[0041] TP (True Positive): The number of image patches that are detected as defective and actually have defects;

[0042] FN (false negative): the number of image blocks that actually have defects but are not detected;

[0043] FP (False Positive): The number of image blocks that are actually defective but are mistakenly judged to be defective;

[0044] TN (True Negatives): The number of image blocks that are actually free of defects and are not mistakenly judged as defective.

[0045] The beneficial effects of the present invention are:

[0046] The present invention provides a method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm.

[0047] 1. Incorporating an adaptive enhancement algorithm, it can dynamically correct missed and misjudged samples that occur during the detection process: Traditional defect detection methods typically rely on static models or a single training data set, ignoring difficult-to-detect sample features, which affects the accuracy of the model. This invention, however, introduces an adaptive enhancement algorithm that iteratively adjusts sample weights based on detection results. In particular, it assigns higher extraction weights to misjudged or missed samples, prompting the model to pay more attention to these difficult-to-identify defects in subsequent training. This processing approach can effectively enhance the model's ability to detect complex defects, significantly improve detection accuracy, and adapt to ever-changing detection scenarios.

[0048] 2. The strategy of integrating multiple base models is adopted to provide more stable and efficient detection results: Unlike traditional methods that rely on the output results of a single model, the present invention generates multiple base models through multiple iterative training. Each base model has different advantages under different detection conditions and scenarios. Finally, an integrated voting mechanism is adopted to output the final defect judgment result based on the judgment results of the majority of base models. This strategy not only improves the recall rate of defect detection, but also effectively balances the false positive rate, avoiding the high false positive rate problem caused by the limitations of a single model in traditional methods, and by dynamically adjusting the threshold of integrated voting, the final judgment is made more accurate, ensuring the high accuracy and reliability of the detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The above or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0050] Figure 1 : Flowchart of the method for detecting metal surface defects of automobile interior parts based on the adaptive enhancement algorithm of the present invention;

[0051] Figure 2 : The training process of the base model of the present invention;

[0052] Figure 3 : Example diagram of image annotation software;

[0053] Figure 4 : Example diagram of defect detection result visualization. DETAILED DESCRIPTION

[0054] The present invention is described below in conjunction with specific embodiments:

[0055] See also Figure 1-4 The present invention proposes a metal surface defect detection method for automobile interior parts based on an adaptive enhancement algorithm. First, the physical properties of the metal product surface to be tested and the type of defects to be tested are determined according to the actual application background, and the real metal workpiece surface image is obtained and a data set is generated; then, according to the following Figure 2 The training process of the base model shown in the figure uses the prepared data set to iteratively train several base models; finally, in the actual defect detection process, the detection results of these base models are integrated to obtain the final result. The specific steps of implementing the present invention include:

[0056] Step 1: Acquire images;

[0057] Acquire images of real metal workpiece surfaces similar to the metal product to be tested, which can be obtained from public datasets or captured by a custom-built imaging system.

[0058] Step 2: Data cleaning;

[0059] Since the camera field of view is required to be larger than the range of the metal workpiece during shooting, the unprocessed image often has parts that do not need to be paid attention to around it. The part of the image containing the metal workpiece surface is the valid area, so the real metal workpiece surface image needs to be preprocessed: first delete the part of the image that does not contain the metal workpiece surface, and then cut the effective area of the real metal workpiece surface image into a size suitable for defect detection, and then perform grayscale space transformation on the metal workpiece surface image to avoid the overall image being too dark or the grayscale value distribution being too concentrated.

[0060] Step 3: Manual labeling;

[0061] Determine the type of defects to be tested and have experts pass the Figure 3 The image annotation software shown in the figure manually annotates the preprocessed real metal workpiece surface image and generates a label file. The label file format is txt and contains the type and location information of the defect. This label file and the real metal workpiece surface image together constitute the real metal workpiece surface image dataset.

[0062] Step 4: Train the model;

[0063] Step 4.1: Each sample in the real metal workpiece surface image dataset corresponds to a weight. The normalized value of the weight represents the probability of the sample being drawn. Before the first base model is trained, each sample is assigned the same initial weight. Subsequent weights may change due to samples being classified as difficult to distinguish.

[0064] Step 4.2: Randomly extract 80% of the total number of real metal workpiece surface image datasets according to weight as the training dataset for the first base model. The detailed process of extracting the training dataset is as follows:

[0065] During extraction, a number is randomly generated from 0 to 1 each time. Each sample corresponds to a range. The range length of a sample is the result of normalizing the sample weight. The ranges of all samples are connected end to end without overlapping, covering all ranges between 0 and 1. When the randomly generated number is within the range corresponding to a sample, the sample is placed in the training data set. The number of samples in the final training data set is equal to the number of times the above extraction is repeated. Since the same sample in the training data set will not be overwritten when the same sample is repeatedly drawn, it is possible that some samples in the training data set appear multiple times.

[0066] In this process, it is necessary to ensure that the number of defective samples and non-defective samples is balanced, that is, the ratio of defective samples to non-defective samples in the extracted training dataset and the metal workpiece surface image dataset matches.

[0067] Step 4.3: Train the base model

[0068] The base model structure is a convolutional neural network. In the deep learning object detection framework, the model generally consists of four parts: input, backbone network, neck, and detection head. Their specific functions are as follows:

[0069] (1) Input

[0070] The main function of the input part is to convert the original image into a data format suitable for model processing. At the same time, it performs necessary data augmentation and preprocessing operations, including resizing the input image to a fixed size (such as 640 pixels × 640 pixels) to meet the input size requirements of the model and normalizing the image pixel values (such as scaling the pixel values to the range [0, 1]). In addition, data augmentation techniques (such as random cropping, flipping, rotation, brightness adjustment, etc.) are also applied at this stage to increase data diversity and help the model improve its robustness to complex scenes.

[0071] (2) Backbone

[0072] The backbone network is the core component of the object detection framework for feature extraction. It is responsible for extracting basic visual features from the input image. Through multi-layer convolution operations, it can capture low-level features (such as texture and edges) and high-level semantic features (such as shape and category information) in the image. Modern object detection frameworks typically introduce optimized architectural designs into the backbone network to balance computational efficiency and feature expression capabilities, providing high-quality features for subsequent object detection tasks.

[0073] (3) Neck

[0074] The neck part is used to further process and integrate the features generated by the backbone network to generate multi-scale feature maps to meet the requirements of target detection for different target sizes. It usually adopts structures such as Feature Pyramid Network (FPN) or Path Aggregation Network (PAN). By fusing multiple layers of features, combining shallow detail information with deep semantic information, it improves the detection ability of small and large targets. This part lays the foundation for multi-scale target detection and adapts to diverse targets in complex scenes.

[0075] (4) Detection head

[0076] The detection head is responsible for predicting the location and category of the target based on the feature map generated by the neck. It is the output module of the entire target detection framework. Through classification and regression operations, the detection head can output the category probability of each target and the coordinate information of the bounding box. Modern detection head design tends to simplify calculations. For example, it uses an Anchor-Free approach to directly predict the target center point and bounding box, thereby improving detection efficiency. This part of the design also supports multi-scale output, enabling the model to have good detection performance for both large and small targets.

[0077] Step 4.4: Defect detection;

[0078] The trained base model is used to detect the real metal workpiece surface image dataset, and the detection results are compared with the expert manual annotation results to screen out missed detection samples and misjudged samples, which are collectively referred to as incorrectly detected samples.

[0079] The definitions of missed samples and misjudged samples are as follows: if the expert does not mark the defect during manual annotation, but the model detects the defect in the sample, then the sample is a misjudged sample; similarly, if the expert marks the defect during manual annotation, but the model does not detect the defect, then the sample is a missed sample.

[0080] Step 4.5: Increase the weight of the false detection sample, and keep the weights of the other samples unchanged. Save the weights of all samples in the real metal workpiece surface image dataset in a table file. The first column of the table is the name of the sample, and the second column is the corresponding weight.

[0081] Step 4.6: Repeat steps 4.2 to 4.5 to iteratively train multiple base models;

[0082] The number of base models depends on the specific situation, but before the first round of training, the weights of all samples are the same. During the iterative training process, if some difficult-to-detect samples are incorrectly detected multiple times by the base model, their weights will become larger and larger, and the probability of being extracted when the training data set is extracted will also become higher and higher. Therefore, in order to avoid the serious imbalance of the extracted training data set due to the excessive weight of some samples after being incorrectly detected multiple times, and the presence of a large number of repeated samples, you can consider setting a weight upper limit threshold as needed. When the weight reaches this threshold, it will no longer increase. This training method can effectively improve the model's attention to difficult-to-detect samples.

[0083] Step 5: Performance testing;

[0084] Due to the randomness introduced by the training process, even if the exact same training data set and hyperparameters are used, the trained models will have individual differences. Therefore, the defect detection performance of each base model should be tested individually, and base models that meet the performance requirements should be screened out, while base models with poor performance should be discarded. The detection results of the remaining base models should be integrated, and the final result should be obtained through an integrated voting mechanism.

[0085] In the above implementation process, the present invention improves the performance of the detection system through the following two key innovations:

[0086] (1) Implementation of the adaptive enhancement algorithm: During the training process, the weight of each sample is adjusted based on its performance in the base model. The weight of missed and misjudged samples will be increased, resulting in a higher probability of these samples being drawn in the next round of training. This adaptive adjustment mechanism can effectively increase the base model's attention to difficult-to-detect samples. By gradually correcting the model's misjudgments and missed detections, the algorithm can make each iteration more accurate, thereby significantly improving the effectiveness of the entire defect detection process.

[0087] (2) Base model integration and voting mechanism: By integrating multiple base models, the integrated model can take advantage of the complementary advantages of multiple base models and provide more stable and efficient detection results compared to the detection results of a single model. Each base model may have better recognition capabilities for different types of defects in different detection tasks. Therefore, by integrating these base models, more detection scenarios can be covered and the recall rate can be further improved. In addition, the integrated voting mechanism can effectively reduce false positives. Only when the majority of base models determine that a sample contains a defect will the final result output a defect. This method significantly reduces the false positive rate and ensures high-quality detection results.

[0088] Table 1 compares the characteristics of the defect detection method proposed in the present invention and the traditional single-model defect detection method.

[0089] Table 1 Comparison between the present invention and the traditional single model defect detection method

[0090]

[0091] The feasibility of the metal surface defect detection method for automotive interior parts based on the adaptive enhancement algorithm proposed in this paper is verified through experiments.

[0092] The dataset used in the experiment was obtained from a domestic company's automotive interior metal sheet materials through push scanning with a high-resolution imaging system. After preprocessing, a total of 12,101 640-pixel × 640-pixel image blocks were obtained. Of these, 6,196 (2,572 with defects and 3,624 without defects) were used to train the base model, and 5,905 (2,507 with defects and 3,398 without defects) were used for the final test. During the test, visual defect detection results were generated, as shown in the figure below. Figure 4 As shown, it includes the metal surface image of the automobile interior decoration parts in the corresponding area, the defect position framed on the image, and the defect type and actual size marked.

[0093] The hardware conditions for training the model are 8 2080Ti servers for parallel training, and a total of 9 base models were trained. When a base model detects defects in the real automobile interior surface image dataset (that is, 6196 image blocks used for training), some non-defective samples are detected as defects or defective samples are not detected as defects, this type of sample is the wrong detection sample in this round of training, and the corresponding weight of this type of sample increases. When the training dataset used for the next base model is extracted, the probability of being extracted will increase. Moreover, in the process of iterative training, the sample weight will be inherited, that is, if a sample is incorrectly detected by multiple base models, then the probability of the sample being extracted will become larger and larger, as follows: the initial weight of the sample is 1. Assuming that the sample is incorrectly detected by m base models before a certain extraction, its weight is changed to e^m.

[0094] The evaluation indicators of this experiment mainly include recall rate (Recall) and false positive rate (FPR), as follows:

[0095]

[0096] TP (True Positives): The number of image patches that are detected as having defects and actually have defects.

[0097] FN (False Negative): The number of image patches that actually have defects but are not detected.

[0098] FP (False Positive): The number of image blocks that are actually free of defects but are mistakenly judged to be defective.

[0099] TN (True Negatives): The number of image blocks that are actually free of defects and are not mistakenly judged as defective.

[0100] The test results are shown in Table 2. The last row is the final output after the ensemble voting of the nine base models. The voting rule is: if there are at least five base models with defects in the predicted sample, it is ultimately judged to be defective; otherwise, it is ultimately judged to be non-defective.

[0101] Table 2 Test results

[0102]

[0103] As can be seen from the table, the defect detection results obtained by this method achieve a good balance between the recall rate and the false positive rate. It will not lead to a decline in product quality due to a low recall rate, nor will it lead to a large amount of raw material waste due to a high false positive rate.

[0104] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.

Claims

1. A method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm, characterized in that: The following steps are involved: Step 1: Acquire images; Acquire the surface image of a real metal workpiece similar to the metal product to be tested; Step 2: Data cleaning; Step 3: Manual labeling; Determine the type of defects to be detected and manually annotate the pre-processed real metal workpiece surface images to generate a real metal workpiece surface image dataset; Step 4: Train the model; Step 4.1: Assign the same initial weight to each sample in the real metal workpiece surface image dataset; Step 4.2: Randomly extract 80% of the total number of real metal workpiece surface image datasets according to weights as the training dataset for the first base model; Step 4.3: Train the base model; Step 4.4: Defect detection; Use the trained base model to detect real metal workpiece surface image datasets, and compare the detection results with the expert manual annotation results to filter out incorrect detection samples; Step 4.5: Increase the weight of the falsely detected samples, and keep the weights of the remaining samples unchanged; Step 4.6: Repeat steps 4.2 to 4.5 to iteratively train multiple base models. The number of base models depends on the specific situation. Step 5: Performance testing; The defect detection performance of each base model is tested individually, and the base models that meet the performance standards are screened out, while the base models with poor performance are discarded; the detection results of the remaining base models are integrated, and the final result is obtained through an integrated voting mechanism.

2. The method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm according to claim 1, characterized in that: The step 2 is to pre-process the real metal workpiece surface image obtained in step 1, specifically including: First, the part of the image that does not contain the metal workpiece surface is deleted, and then the remaining part is cut into a size suitable for defect detection. Then, the grayscale space transformation of the metal workpiece surface image is performed to avoid the overall image being too dark or the grayscale value distribution being too concentrated.

3. The method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm according to claim 1, characterized in that: The real metal workpiece surface image dataset generated in step 3 includes the type and location information of the defects and the real metal workpiece surface image.

4. The method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm according to claim 1, characterized in that: In step 4.2, the number of defective samples and non-defective samples is balanced when extracting the training data set. The detailed process is as follows: During extraction, a number is randomly generated from 0 to 1 each time. Each sample corresponds to a range. The range length of a sample is the result of normalizing the sample weight. The ranges of all samples are connected end to end without overlapping, covering all ranges between 0 and 1. When the randomly generated number is within the range corresponding to a sample, the sample is placed in the training data set. The number of samples in the final training data set is equal to the number of times the above extraction is repeated.

5. The method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm according to claim 1, characterized in that: In step 4.3, the base model structure is a convolutional neural network, which consists of input, backbone network, neck and detection head; The input part converts the original image into a data format suitable for base model processing, and performs necessary data enhancement and preprocessing operations; The backbone network is responsible for extracting basic visual features from the input image and capturing low-level features and high-level semantic features in the image through multi-layer convolution operations. The neck part further processes and integrates the features generated by the backbone network to generate multi-scale feature maps to meet the requirements of target detection for different target sizes; The detection head part predicts the location and category of the target based on the feature map generated by the neck, and outputs the category probability of each target and the coordinate information of the bounding box.

6. The method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm according to claim 1, characterized in that: In step 4.4, the error detection samples include missed detection samples and misjudgment samples; The missed detection samples are samples that were marked as defects by experts during manual annotation, but were not detected by the model; The misjudged samples are samples in which defects are not marked when the experts manually annotate them, but the model detects defects in the samples.

7. The method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm according to claim 1, characterized in that: In step 4.5, when the weight reaches an upper threshold, it will no longer increase, so as to avoid the weight of some samples being too large after being wrongly detected multiple times, which will cause serious imbalance in the training data set after extraction.

8. The method for detecting metal surface defects of automobile interior parts based on an adaptive enhancement algorithm according to claim 1, characterized in that: The indicators for evaluating the defect detection performance of the base model in step 5 mainly include the recall rate Recall and the false positive rate FPR, which are as follows: Where, TP (True Positive): The number of image patches that are detected as defective and actually have defects; FN (false negative): the number of image blocks that actually have defects but are not detected; FP (False Positive): The number of image blocks that are actually defective but are mistakenly judged to be defective; TN (True Negatives): The number of image blocks that are actually free of defects and are not mistakenly judged as defective.

Images