Propeller blade surface damage intelligent inspection method and system based on machine vision

Abstract

The invention relates to the technical field of machine vision, and discloses a propeller blade surface damage intelligent inspection method and system based on machine vision, and the method comprises the steps: enabling an annular LED light source array and an image collection device to be coaxially arranged, and enabling the annular LED light source array and the image collection device to be aligned with a to-be-inspected region; according to the pre-estimated three-dimensional curved surface model of the to-be-detected area, sequentially lightening the illumination units in the light source array according to an optimized and determined sequence, and synchronously acquiring single-angle images to form an image set; performing damage feature extraction on each image to obtain a damage feature map; calculating a pixel-level weight map corresponding to each single-angle image based on all the damage feature maps; pixel-level weighted fusion is carried out on all the single-angle images according to the weight map, and a fused image is generated; the fused image is analyzed to identify and classify surface damage. Through intelligent illumination optimization and feature adaptive image fusion based on the curved surface model, illumination interference caused by a complex curved surface is effectively overcome, and reliable detection of propeller blade surface damage is realized.

Description

Technical Field

[0001] This application relates to the field of machine vision technology, and in particular to a machine vision-based intelligent inspection method and system for surface damage of propeller blades. Background Technology

[0002] As a critical power component, ship propellers are exposed to the complex and harsh marine environment for extended periods. Their blade surfaces are prone to cracks, corrosion, and other damage. If these minor damages are not detected in time, they can lead to structural failure and seriously threaten navigation safety. Currently, propeller surface inspection mainly relies on manual visual inspection or handheld inspection equipment. This method is inefficient, subjective, and prone to missing defects, making it difficult to meet the needs of large-scale, high-standard inspections.

[0003] With the development of machine vision technology, image-based automatic inspection methods have become a research hotspot. However, applying machine vision to workpieces with complex spatial curved surfaces, such as propellers, faces significant challenges. The most significant difficulty lies in controlling the lighting conditions. Due to the varying normal directions at different points on the curved surface, under a fixed single light source, strong specular highlights are easily formed in some areas, while deep shadows are produced in other areas. Highlights can obscure damage details and cause localized overexposure of the image; shadows can reduce the contrast between the damage and the background, or even completely hide features. This uneven lighting problem caused by the curved surface geometry seriously interferes with the stable extraction and recognition of damage features by subsequent image processing algorithms, resulting in insufficient reliability, high false negative and false positive rates of traditional visual inspection methods on curved workpieces. Therefore, designing a visual inspection method that can adapt to complex curved surface geometry and actively overcome lighting interference is key to achieving intelligent and reliable inspection of propeller blade surface damage. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of this application provide a machine vision-based intelligent inspection method for propeller blade surface damage to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the machine vision-based intelligent inspection method for propeller blade surface damage provided in this application includes: S1. Set the ring-shaped LED light source array coaxially with the image acquisition device and align it with the area to be inspected on the propeller blades; S2. Based on the estimated three-dimensional surface model of the area to be inspected, each illumination unit in the ring LED light source array is lit sequentially in the optimized order, and each time it is lit, the image acquisition device simultaneously acquires a single-angle image of the area to be inspected, thereby obtaining a set of single-angle images. S3. Extract damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image; S4. Based on each of the damage feature maps, calculate the pixel-level weight map corresponding to each image in the single-angle image set; S5. Based on the pixel-level weight maps, perform pixel-level fusion on all images in the single-angle image set to generate a fused image for damage recognition; S6. Analyze the fused image to identify surface damage to the propeller blades.

[0006] 2. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, the step of coaxially setting the annular LED light source array with the image acquisition device and aligning it with the area to be inspected on the propeller blade specifically includes: The ring-shaped LED light source array is divided into multiple independent lighting sectors; Control multiple lighting sectors so that the lighting sectors are individually lit sequentially along a circular path around the optical axis of the image acquisition device; When each of the lighting sectors is illuminated, the image acquisition device is controlled to acquire a single-angle image of the area to be inspected. The single-angle images corresponding to all the lighting sectors are collected to form the single-angle image set.

[0007] 3. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, according to the estimated three-dimensional surface model of the area to be inspected, each illumination unit in the annular LED light source array is sequentially illuminated in an optimized order, and each time it is illuminated, the image acquisition device synchronously acquires a single-angle image of the area to be inspected, thereby obtaining a set of single-angle images, specifically including: Based on the estimated three-dimensional surface model of the area to be inspected, the relationship between the illumination direction of each lighting unit in the ring LED light source array and the local normal vector of each point on the estimated three-dimensional surface model is analyzed to determine the lighting order of the lighting units. The lighting order of the lighting units is configured such that for any point in the area to be inspected, at least one lighting unit has an angle greater than 60 degrees between its illumination direction and the local normal vector of that point under the lighting order. Based on the lighting sequence of the lighting units, a lighting control command sequence corresponding one-to-one with each lighting unit in the ring LED light source array is generated; According to the lighting control command sequence, the corresponding lighting units are driven to emit light in sequence, and a synchronous trigger signal is sent to the image acquisition device at the same time; The image acquisition device responds to the synchronization trigger signal, acquires and outputs a single-angle image corresponding to the currently lit lighting unit; According to the driving order of the lighting units, the sequentially output single-angle images are sorted and cached to form a structured set of single-angle images.

[0008] 4. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, the step of extracting damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image specifically includes: Each image in the single-angle image set is subjected to image enhancement processing to obtain an enhanced image; The enhanced image is subjected to edge detection processing to obtain the corresponding edge response map; The edge response map is subjected to binarization segmentation to obtain a preliminary segmented binary image; Morphological filtering is performed on the initially segmented binary image to eliminate isolated pixel regions with non-damaging features, resulting in a basic damage feature map. For each connected region in the basic damage feature map, the region is classified based on its morphological feature parameters, and the region is marked as a linear crack feature region or a planar corrosion feature region, thereby generating a damage feature map containing damage type information.

[0009] 5. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 4, characterized in that, marking the connected region as a linear crack feature region or a planar corrosion feature region, and generating a damage feature map containing damage type information, specifically includes: Read the information of the pre-marked linear crack feature region and planar corrosion feature region in the damage feature map; A base weight gain coefficient is assigned to the pixels belonging to the linear crack feature region, which is greater than the base weight gain coefficient used by the pixels in the planar corrosion feature region.

[0010] 6. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, the step of calculating the pixel-level weight map corresponding to each image in the single-angle image set based on each of the damage feature maps specifically includes: The damage feature maps are smoothed by a smoothing filter to obtain smoothed feature maps. Based on the smoothed feature values ​​of all images at the same pixel position in the single-angle image set, normalization calculation is performed on each smoothed feature map to obtain an initial weight map. The initial weight map is constrained to ensure that the initial weight values ​​corresponding to each pixel position are non-negative and their sum is 1, thus obtaining the pixel-level weight map.

[0011] 7. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, the step of performing pixel-level fusion on all images in the single-angle image set according to each pixel-level weight map to generate a fused image for damage identification specifically includes: S51. Obtain each single-angle image in the single-angle image set and the pixel-level weight map corresponding to each single-angle image; S52. For each pixel position, the pixel values ​​of all the single-angle images corresponding to that position are weighted and calculated with the weight value of that position in their respective pixel-level weight map, and the weighted results are summed to obtain the pixel value of that position in the fused image. S53. Repeat the weighted summation of step S52 for all pixel locations to generate the fused image.

[0012] 8. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that the step of analyzing the fused image to identify the surface damage of the propeller blade specifically includes: The fused image is subjected to damage feature enhancement and segmentation processing to obtain a binary mask image containing candidate damage regions; Based on predefined damage morphology and grayscale features, each candidate damage region in the binary mask image is verified and classified to determine whether the candidate damage region is a real damage and the damage type. The location and type information of the verified and classified damage areas are integrated and output to form the final damage identification result.

[0013] 9. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 8, characterized in that, the step of verifying and classifying each candidate damage region in the binary mask image based on predefined damage morphology and grayscale features, and determining whether the candidate damage region is a real damage and the damage type, specifically includes: Extract the morphological feature parameters and grayscale statistical feature parameters of the corresponding region in the fused image for each candidate damage region in the binary mask image; The morphological feature parameters and grayscale statistical feature parameters of the candidate damage region are compared with a preset damage feature threshold range. If they all fall within the corresponding preset damage feature threshold range, they are determined to be real damage. For areas determined to be real damage, they are further classified as crack damage or corrosion damage based on the morphological characteristic parameters.

[0014] To address the aforementioned problems, this application also provides a machine vision-based intelligent inspection method system for propeller blade surface damage, the system comprising: The coaxial illumination and image acquisition module is used to coaxially set the ring LED light source array with the image acquisition device and align it with the area to be inspected on the propeller blades. The model-driven lighting optimization acquisition module is used to sequentially illuminate each lighting unit in the ring LED light source array according to the estimated three-dimensional surface model of the area to be inspected, and to simultaneously acquire a single-angle image of the area to be inspected by the image acquisition device each time it is illuminated, thereby obtaining a set of single-angle images. The damage feature extraction and classification module is used to extract damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image. An adaptive weight map calculation module is used to calculate the pixel-level weight map corresponding to each image in the single-angle image set based on each of the damage feature maps. A pixel-level image fusion module is used to perform pixel-level fusion on all images in the single-angle image set according to the pixel-level weight maps to generate a fused image for damage recognition. The damage intelligent recognition module is used to analyze the fused image to identify surface damage to the propeller blades.

[0015] Compared with existing technologies, this application brings significant technological advancements by introducing an active illumination optimization and feature adaptive fusion mechanism based on prior geometric knowledge: First, this method fundamentally optimizes the information quality at the image acquisition source. By intelligently planning the illumination sequence based on the 3D model of the surface to be inspected, it ensures that every point on the surface can be imaged under at least one grazing illumination condition, thereby proactively creating the most favorable illumination environment for highlighting the minute uneven features of the surface. This directional illumination strategy effectively suppresses specular highlights and enhances the imaging contrast of the damaged area from the data acquisition stage, providing a high-quality original image sequence rich in damage information for subsequent processing.

[0016] Secondly, this method achieves intelligent synthesis from multi-angle image information to a single optimized view. By extracting damage features from each image and calculating pixel-level fusion weights accordingly, the system can adaptively evaluate the contribution value of each image at each local location and perform weighted fusion. This process is not a simple image overlay, but a data-driven decision based on the saliency of damage features. It effectively integrates the advantageous information under different lighting angles, ultimately generating a fused image with globally enhanced damage features and maximally suppressed background interference. Based on this, combined with quantitative verification and classification of morphological and grayscale features, accurate and automated damage identification and classification are achieved. The entire scheme is interconnected, forming a complete technical closed loop from intelligent lighting, feature extraction, adaptive fusion to automatic identification, significantly improving the accuracy, robustness, and automation level of damage detection under complex curved surface conditions. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a machine vision-based intelligent inspection method for surface damage of propeller blades provided in an embodiment of this application. Figure 2 A functional block diagram of a machine vision-based intelligent inspection method system for propeller blade surface damage provided in an embodiment of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0019] This application provides a machine vision-based intelligent inspection method for propeller blade surface damage. The executing entity of this machine vision-based intelligent inspection method for propeller blade surface damage includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the machine vision-based intelligent inspection method for propeller blade surface damage can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating an intelligent inspection method for propeller blade surface damage based on machine vision, provided in an embodiment of this application.

[0021] In the embodiments of this application, the reliable detection of surface damage (such as cracks and corrosion) of ship propellers, a key component with complex spatial curved surfaces, is a core technical problem to ensure navigation safety. Traditional visual methods are difficult to overcome the interference of highlights and shadows caused by curved surfaces under single illumination, resulting in missed or false detection of damage.

[0022] This application provides a machine vision-based intelligent inspection method for surface damage on propeller blades, including: S1. Set the ring-shaped LED light source array coaxially with the image acquisition device and align it with the area to be inspected on the propeller blades.

[0023] In the embodiments of this application, step S1 and its specific implementation aim to construct a basic image acquisition system that can actively and orderly generate multi-angle lighting conditions. This is the primary physical basis for solving the core technical problem that the complex curved surface of the propeller causes low contrast of damage features under single illumination and is easily covered by highlights or shadows. This step is not a simple static photo, but rather generates original, structured multi-angle image data sources for all subsequent advanced image processing through the coordination of hardware structure and control logic.

[0024] In some embodiments, the step of coaxially aligning the annular LED light source array with the image acquisition device and aiming it at the area to be inspected on the propeller blades specifically includes: The ring-shaped LED light source array is divided into multiple independent lighting sectors; Control multiple lighting sectors so that the lighting sectors are individually lit sequentially along a circular path around the optical axis of the image acquisition device; When each of the lighting sectors is illuminated, the image acquisition device is controlled to acquire a single-angle image of the area to be inspected. The single-angle images corresponding to all the lighting sectors are collected to form the single-angle image set.

[0025] In this embodiment, the ring LED light source array is an illumination device composed of multiple light-emitting diode units arranged in a ring structure; the image acquisition device is an industrial camera with an image sensor and an optical lens; coaxial arrangement refers to an installation structure in which the physical central axis of the ring LED light source array coincides with the optical axis of the optical lens of the image acquisition device; the area to be inspected is the surface portion on the propeller blade designated for damage detection; the illumination sector is an arc-shaped light-emitting segment in the ring LED light source array that is divided into independently controllable power-on states; the optical axis of the image acquisition device is the theoretical axis passing through the center of the industrial camera lens and perpendicular to the camera's imaging sensor; a single-angle image is a two-dimensional digital image of the area to be inspected captured by the image acquisition device under the illumination conditions provided by a specific single illumination sector; a single-angle image set is an image sequence or image group composed of all single-angle images acquired in a specific order.

[0026] First, during the hardware assembly phase, the operator secures the ring-shaped LED light source array in front of the industrial camera lens using a mechanical bracket. Fine-tuning is then performed using calibration tools to ensure the geometric center of the light source ring aligns with the physical center of the camera lens, thus achieving optical axis coincidence. This arrangement ensures that the illumination light is projected onto the area to be inspected along the direction near the center of the camera's field of view. Subsequently, the system logically divides the ring-shaped LED light source array, for example, dividing a 360-degree ring light source into 12 independent illumination sectors. Each sector corresponds to a 30-degree arc-shaped light-emitting area, and all LED units within each sector are driven by an independent control circuit. When the system initiates the inspection process for a specific area, the control unit (such as a PLC or embedded microcontroller) begins executing the preset control program.

[0027] In this embodiment of the application, multiple lighting sectors are controlled so that the lighting sectors are lit up sequentially and individually according to a circular path around the optical axis of the image acquisition device. The specific technical means is that the control unit sends a series of pulse width modulation signals with strict timing relationship to the driving circuit.

[0028] For example, the control unit first sends a high-level signal to the drive circuit connected to the first sector, causing the LED in that sector to receive operating current and light up, while the drive circuits of the other sectors remain low-level and off. After a predetermined time sufficient for the camera to complete one exposure (e.g., 100 milliseconds), the control unit switches the control signal of the first sector low to turn it off, and immediately switches the control signal of the second sector high to turn it on. This process is repeated sequentially along a circular path, for example, clockwise, until all N sectors have been lit in turn.

[0029] In this embodiment, when each illumination sector is lit, the image acquisition device acquires a single-angle image of the area to be inspected. This relies on strict hardware synchronization. After each switch and stable illumination of an illumination sector, the control unit simultaneously sends a trigger pulse signal to the industrial camera. Upon receiving this synchronization trigger signal, the camera immediately performs an image acquisition, exposing the area to be inspected currently illuminated by the single active illumination sector and converting it into digital image data. Since the illumination conditions are singular and changeable during each acquisition, each image records the surface reflection of the area to be inspected under light from a specific direction.

[0030] In this embodiment, the single-angle images corresponding to all the illumination sectors are aggregated to form the single-angle image set, which is accomplished through data stream management. After each acquisition, the industrial camera transmits the generated single-angle image data to a host computer or storage buffer via a data interface (such as GigEVision or USB3Vision). The host computer software or control program indexes and arranges these images according to the acquisition sequence number or timestamp attached to the images, in the order in which the sectors are lit (such as from sector 1 to sector N), and stores this ordered set of images in a specified data structure (such as a list or array). This structured dataset is the single-angle image set, which provides a clear and complete raw input for subsequent processing.

[0031] In this embodiment of the application, by sequentially illuminating illumination sectors at different locations, the method simulates the effect of light illuminating a curved surface from different annular directions under a fixed camera viewpoint. Thus, in a single inspection process, image representations of the same area under multiple illumination angles are obtained. This directly overcomes the limitation that a single fixed light source cannot adapt to different orientations of various points on the curved surface, and increases the possibility of revealing damage features hidden by shadows or highlights from the data source.

[0032] S2. Based on the estimated three-dimensional surface model of the area to be inspected, each illumination unit in the ring LED light source array is lit sequentially in the optimized order, and each time it is lit, the image acquisition device simultaneously acquires a single-angle image of the area to be inspected, thereby obtaining a set of single-angle images.

[0033] In this embodiment, step S2 addresses the core technical problem of how to proactively and intelligently generate optimal lighting conditions to maximize the visibility of damage features from the source when performing damage detection on complex curved surfaces (such as propeller blades). This step goes beyond simple multi-angle image acquisition. By introducing prior geometric knowledge of the object under inspection (predicted three-dimensional curved surface model), the lighting sequence is optimized in a directional manner, thereby ensuring that the acquired image sequence itself is rich in high-contrast damage information.

[0034] In some embodiments, the step of sequentially illuminating each illumination unit in the annular LED light source array according to an optimized order based on the estimated three-dimensional surface model of the area to be inspected, and simultaneously acquiring a single-angle image of the area to be inspected by the image acquisition device each time it is illuminated, thereby obtaining a set of single-angle images, specifically includes: Based on the estimated three-dimensional surface model of the area to be inspected, the relationship between the illumination direction of each lighting unit in the ring LED light source array and the local normal vector of each point on the estimated three-dimensional surface model is analyzed to determine the lighting order of the lighting units. The lighting order of the lighting units is configured such that for any point in the area to be inspected, at least one lighting unit has an angle greater than 60 degrees between its illumination direction and the local normal vector of that point under the lighting order. Based on the lighting sequence of the lighting units, a lighting control command sequence corresponding one-to-one with each lighting unit in the ring LED light source array is generated; According to the lighting control command sequence, the corresponding lighting units are driven to emit light in sequence, and a synchronous trigger signal is sent to the image acquisition device at the same time; The image acquisition device responds to the synchronization trigger signal, acquires and outputs a single-angle image corresponding to the currently lit lighting unit; According to the driving order of the lighting units, the sequentially output single-angle images are sorted and cached to form a structured set of single-angle images.

[0035] The estimated three-dimensional surface model is a digital model used to characterize the spatial shape of the surface of the area to be inspected, obtained through a three-dimensional scanning device or imported from computer-aided design drawings of the propeller.

[0036] In this embodiment, the lighting unit is a single light-emitting diode or a group of closely arranged light-emitting diodes that can be independently controlled to turn on and off, forming a ring-shaped LED light source array; the lighting sequence of the lighting units is a sequence of instructions calculated by the control unit to specify the sequential lighting order of each lighting unit, and this order is not a fixed ring-shaped sequence; the illumination direction is the vector direction from the light-emitting center of the lighting unit to the center point of the area to be inspected; the local normal vector is a unit vector perpendicular to the tangent plane of any sampling point on the estimated three-dimensional curved surface model; the lighting control instruction sequence is a specification of the lighting sequence of the lighting units, and is a set of executable instructions containing the lighting unit number, lighting duration, and timing information; the synchronous trigger signal is an electronic pulse signal sent by the control unit to the image acquisition device to trigger it to immediately acquire an image while driving a specific lighting unit to light up.

[0037] In this embodiment, the system loads the estimated three-dimensional surface model of the area to be inspected during the offline or initialization phase. The processing program in the control unit analyzes the model. Specifically, it traverses a large number of sampling points on the surface of the model and calculates the local normal vector of each sampling point. At the same time, the program calculates the illumination direction vector of each lighting unit in the array relative to the area to be inspected based on the physical installation parameters of the ring LED light source array. Then, the system performs key optimization calculations to determine the lighting sequence of the lighting units.

[0038] In this embodiment of the application, the lighting order of the lighting units is configured such that for any point in the area to be inspected, at least one lighting unit has an angle greater than 60 degrees between its illumination direction and the local normal vector of that point. The realization of this technical requirement relies on an optimization algorithm that uses all lighting units as a candidate set and aims to find the shortest possible subset of lighting units and their arrangement order, so that the aforementioned angle condition can be satisfied for all sampling points on the surface model.

[0039] For example, a greedy algorithm can be used: the algorithm first selects a lighting unit whose illumination direction forms an angle greater than 60 degrees with the maximum number of sampling points on the surface; then, among the remaining unsatisfied sampling points, the next lighting unit that can cover the most of these points is selected; and so on, until all sampling points are covered by the illumination direction of at least one selected lighting unit at an angle greater than 60 degrees. The sequence of selected lighting units and their arrangement is the optimized lighting unit lighting order.

[0040] In this embodiment, a lighting control command sequence corresponding one-to-one with each lighting unit in the ring LED light source array is generated based on the lighting unit lighting sequence. This is the process of converting the optimization result into executable commands. The control unit generates a list of instructions containing the specific lighting unit address code, lighting duration, and sequence index based on the above sequence. For example, the generated sequence may not be a simple clockwise order, but a specific order such as [unit 8, unit 3, unit 12, unit 5].

[0041] In this embodiment, according to the lighting control instruction sequence, the corresponding lighting units are driven to emit light in sequence, and a synchronization trigger signal is sent to the image acquisition device at the same time. This is precise synchronization control at the hardware execution level. The control unit (such as a programmable logic controller) reads the first instruction sequence and sends a high-level signal to the circuit of the corresponding unit 8 in the ring light source driver through its digital output port, so that the LED of unit 8 is lit. At almost the same time, the control unit sends a rising edge pulse to the industrial camera as a synchronization trigger signal through another dedicated trigger line.

[0042] In this embodiment, the image acquisition device, in response to the synchronization trigger signal, acquires and outputs a single-angle image corresponding to the currently illuminated lighting unit. This ensures a strict correspondence between the image and the lighting state. Upon receiving the trigger signal, the industrial camera hardware immediately initiates an exposure. Since only lighting unit 8 is emitting light at this time, the acquired image fully reflects the surface reflection characteristics of the area to be inspected under the specific illumination direction of unit 8. After the preset exposure time, the control unit turns off unit 8 and immediately executes the next instruction in the instruction sequence, driving unit 3 to emit light and synchronously triggering the camera again, thus cycling through the sequence.

[0043] In this embodiment, the sequentially output single-angle images are sorted and cached according to the driving order of the lighting units to form a structured set of single-angle images, thus completing the encapsulation of data into structured information. The host computer software associates the images with the lighting unit numbers in the driving instruction sequence according to the order in which the images arrive or the embedded sequence number, and arranges and stores them according to the driving order (e.g., [image_unit 8, image_unit 3, image_unit 12, image_unit 5]) to form an ordered and indexable set of image data.

[0044] In this embodiment of the application, by optimizing the lighting based on the surface geometry, the system actively creates conditions for each point on the surface to be illuminated at least once by "grazing light" (large incident angle). This lighting can greatly highlight the shadows generated by the micro-unevenness of the surface (such as cracks and pits), thereby significantly improving the contrast between the damage features and the background during the acquisition stage, and effectively suppressing the masking of damage information by overexposure highlights caused by the specular reflection of the curved surface.

[0045] S3. Extract damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image.

[0046] In this embodiment, step S3 aims to automatically identify, segment, and preliminarily classify potential damage features from each image, thereby transforming the original image data into a feature representation rich in semantic information. This step solves the problem of how to stably extract subtle and contrast-varying damage features from multi-angle illumination images, providing accurate and categorized input for subsequent adaptive fusion and final recognition.

[0047] In some embodiments, the step of extracting damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image specifically includes: Each image in the single-angle image set is subjected to image enhancement processing to obtain an enhanced image; The enhanced image is subjected to edge detection processing to obtain the corresponding edge response map; The edge response map is subjected to binarization segmentation to obtain a preliminary segmented binary image; Morphological filtering is performed on the initially segmented binary image to eliminate isolated pixel regions with non-damaging features, resulting in a basic damage feature map. For each connected region in the basic damage feature map, the region is classified based on its morphological feature parameters, and the region is marked as a linear crack feature region or a planar corrosion feature region, thereby generating a damage feature map containing damage type information.

[0048] In some embodiments, marking the connected regions as linear crack feature regions or planar corrosion feature regions and generating a damage feature map containing damage type information specifically includes: Read the information of the pre-marked linear crack feature region and planar corrosion feature region in the damage feature map; A base weight gain coefficient is assigned to the pixels belonging to the linear crack feature region, which is greater than the base weight gain coefficient used by the pixels in the planar corrosion feature region.

[0049] In this embodiment, the damage feature map is an image with the same spatial resolution as the original single-angle image, and its pixel values ​​are used to characterize the probability of damage at the corresponding location or the damage category to which it belongs; the enhanced image is an image processed by contrast or brightness adjustment algorithms, the purpose of which is to improve the visual effect to facilitate feature extraction; the edge response map is an image obtained by calculating the gray-level gradient of the image through convolution operation, where the intensity value of each pixel represents the probability of an edge existing at that location; the binary image is an image containing only two pixel values ​​(usually 0 and 255, representing the background and foreground).

[0050] In this embodiment, morphological filtering is a shape-based image processing technique that filters out noise or connects broken regions by having structural elements interact with the image. A connected region refers to a continuous region in a binary image composed of adjacent pixels of the same type (such as white foreground pixels). Morphological feature parameters are quantitative indicators used to describe the geometric characteristics of a connected region.

[0051] In this embodiment of the application, each image in the single-angle image set is subjected to image enhancement processing. The system reads the single-angle images one by one and operates, for example, by using a histogram equalization algorithm. This algorithm redistributes the pixel intensity values ​​of the image, so that the pixels gathered in a certain gray range are expanded to a wider range, thereby increasing the global contrast of the image and making the difference between the damaged area and the background more obvious. The output result is the enhanced image.

[0052] In this embodiment of the application, edge detection processing is performed on the enhanced image. The system applies the Canny edge detection operator to each enhanced image. The operator first uses a Gaussian filter to smooth the image to reduce noise, then calculates the intensity gradient magnitude and direction of the image, then applies non-maximum suppression to refine the edges, and finally obtains clear boundary lines by double threshold detection and edge connection. The output of this process is the edge response map, in which bright lines represent the detected edges.

[0053] In this embodiment, the edge response map is binarized for segmentation. The system sets a global threshold T_edge for the edge response map, sets all pixels with gradient magnitudes greater than T_edge to 255 (white, representing feature edges), and sets the rest to 0 (black, representing background), thereby obtaining a preliminary segmented binary image containing only the contours of potential damage. Due to noise and texture interference, this binary image usually contains a large number of isolated scattered points or small burrs, so morphological filtering is required.

[0054] In this embodiment, the system first performs an opening operation on the binary image, that is, first performs an erosion operation to remove small isolated noise points, and then performs a dilation operation to restore the approximate size of the remaining area; then, it performs a closing operation, that is, first dilates to connect adjacent but broken short lines, and then performs erosion to smooth the boundaries. After such a combination of opening and closing operations, most of the isolated pixel regions of non-damaging features are eliminated, while the connectivity of real crack or eroded areas is preserved, and the output result is a clear basic damage feature map.

[0055] In this embodiment, each connected region in the basic damage feature map is classified based on its morphological feature parameters. The system uses a region labeling algorithm (such as a two-pass scanning method) to identify all connected regions in the basic damage feature map. For each connected region, its key morphological feature parameters are calculated, including the region area A (total number of pixels), the minimum bounding rectangle length L and width W, and the aspect ratio R (R=L / W, when L≥W). A classification threshold R_thresh is set. If the aspect ratio R of a connected region is greater than R_thresh and its width W is less than a maximum crack width threshold W_crack, then the region is marked as a linear crack feature region; conversely, if the region area A is greater than a minimum corrosion area threshold A_corrosion and its shape is relatively complex (e.g., determined by calculating its roundness or rectangularity), then the region is marked as a planar corrosion feature region. The system generates a new image in which pixels marked as linear crack feature regions are assigned a label value (e.g., 1), pixels marked as planar corrosion feature regions are assigned another label value (e.g., 2), and the background is 0, thus generating a damage feature map containing damage type information.

[0056] In this embodiment of the application, through a standardized image processing workflow, this step can stably extract edge and region information related to damage from each single-angle image and effectively suppress noise interference. At the same time, it can make preliminary distinctions between damage types through morphological analysis, which greatly compresses the amount of data and transforms visual information into a structured feature map with preliminary semantic labels that is more suitable for further computer analysis.

[0057] In this embodiment, the further specific implementation of marking the connected regions as linear crack feature regions or planar corrosion feature regions to generate a damage feature map containing damage type information reflects the deep integration of the method in information flow. This approach not only completes classification but also transforms the classification results into control parameters that influence subsequent fusion strategies. Specifically, the implementation includes: first, the system reads the pre-labeled linear crack feature region and planar corrosion feature region information from the damage feature map, i.e., obtaining the category label for each pixel; then, it assigns a basic weight gain coefficient K1 to pixels belonging to the linear crack feature region, which is greater than the basic weight gain coefficient K2 used by pixels in the planar corrosion feature region.

[0058] For example, setting K1=1.5 and K2=1.0 means that when calculating the initial weights in S4, for the same pixel, if it is identified as a linear crack feature in a single-angle image, the feature value at that point will be multiplied by the gain coefficient K1 (1.5) before subsequent smoothing and normalization; if it is identified as a planar corrosion feature, it will be multiplied by K2 (1.0). This technique is implemented by generating a corresponding gain coefficient mapping map simultaneously with the final damage feature map generated in step S3. When step S4 needs to read the damage feature map of an image for weight calculation, this gain coefficient mapping map will be read synchronously, and the feature map value will be multiplied by the gain coefficient at the corresponding position during the calculation process. The technical effect of this specific implementation is that it treats damage differently at the feature level based on the physical characteristics of the damage (linear cracks are usually finer, have lower contrast, and require stronger emphasis). This allows illumination angles that are more sensitive to crack features and perform better to receive higher fusion weights during the final image fusion process, thus further highlighting the cracks in the fusion result. The connection between this specific implementation and the overall scheme lies in its creative feedback of the feature classification results of S3 to influence the weight allocation strategy of S4, forming a closed loop from feature extraction to fusion optimization, enhancing the overall method's adaptive detection capability for different types of damage.

[0059] S4. Based on each of the damage feature maps, calculate the pixel-level weight map corresponding to each image in the single-angle image set.

[0060] In this embodiment of the application, step S4 constructs a quantitative mapping mechanism from feature saliency to fusion contribution. This step solves the key technical problem of how to objectively and adaptively evaluate the contribution value of each image to the final damage recognition task at each local location in multi-angle images, thereby providing an accurate quantitative basis for generating the optimal fused image.

[0061] In some embodiments, calculating the pixel-level weight map corresponding to each image in the single-angle image set based on each of the damage feature maps specifically includes: The damage feature maps are smoothed by a smoothing filter to obtain smoothed feature maps. Based on the smoothed feature values ​​of all images at the same pixel position in the single-angle image set, normalization calculation is performed on each smoothed feature map to obtain an initial weight map. The initial weight map is constrained to ensure that the initial weight values ​​corresponding to each pixel position are non-negative and their sum is 1, thus obtaining the pixel-level weight map.

[0062] In this embodiment, the pixel-level weight map is a digital image with the same size as the single-angle image. The weight value stored at each pixel position is a real number between 0 and 1, which determines the contribution ratio of the corresponding single-angle image to the pixel value of the final fused image at the corresponding pixel position. Smoothing filtering is an image processing technique used to reduce random noise or local outliers in an image. It is achieved through weighted averaging within the pixel neighborhood. The smoothed feature map is an image obtained after smoothing filtering of the damaged feature map, resulting in a more gradual change in feature response values. The initial weight map is a weight distribution map that initially reflects the relative importance of each image after normalization calculation, but its values ​​may not yet fully satisfy the probability distribution constraints. Constraint processing is a calculation process that adjusts a set of values ​​to force them to meet specific mathematical conditions.

[0063] In this embodiment of the application, the system performs smoothing filtering on each of the damage feature maps. Since the damage feature maps generated in step S3 may contain small abrupt changes due to image noise or inaccurate segmentation, directly using them to calculate weights would lead to instability of the weight map. Therefore, the system performs Gaussian filtering on each damage feature map independently.

[0064] Specifically, for a pixel location (x,y), its smoothed feature value F'_n(x,y) is calculated by weighted averaging of its neighboring pixels. The weights are determined by a two-dimensional Gaussian function G(u,v). After this processing, a smoothed feature map is obtained, which effectively suppresses the interference of local noise on the weight allocation.

[0065] In this embodiment, the system performs a crucial normalization calculation on each smoothed feature map based on the smoothed feature values ​​of all images in the single-angle image set at the same pixel position. For any specific pixel position (x, y) in space, assuming there are N single-angle images, this position corresponds to N smoothed feature values: F'_1(x, y), F'_2(x, y), ..., F'_N(x, y). The system uses the Softmax function as the core means of normalization calculation, calculating an initial weight value W'_n(x, y) for the nth image at this position. By traversing all pixel positions and performing this normalization operation based on the feature values ​​of all images on each smoothed feature map, the system generates an initial weight map for each single-angle image.

[0066] In this embodiment, the system performs constraint processing on the initial weight map. Although the Softmax function theoretically guarantees normalization, in actual numerical calculations, due to floating-point precision limitations or extreme input values, an additional constraint step may be needed to ensure mathematical rigor. The specific technique for this processing is as follows: for each pixel position (x, y), the system reads the values ​​W'_1(x, y)...W'_N(x, y) from all N initial weight maps at that position. First, any extremely small negative values ​​(caused by calculation errors) are set to zero to ensure they are all non-negative; then, their sum is calculated; finally, each value is divided by this sum. This step enforces a hard constraint that the sum of all weight values ​​corresponding to each pixel position is 1. The final weight map obtained after this constraint processing is the pixel-level weight map.

[0067] In this embodiment, this step ensures that image regions with more pronounced damage features under a specific illumination angle receive greater weight during fusion. For example, for a crack, if its features (edge ​​response) are very prominent (large F'_n value) in certain lateral illumination images, the weight W_n of these images in the crack region will be automatically increased through the Softmax function; while in frontal illumination images where the crack is not obvious, its weight is reduced. This essentially allows the image sequence itself to "vote" to determine which image(s) or images(s) information is most worth retaining at each local point.

[0068] S5. Based on the pixel-level weight maps, perform pixel-level fusion on all images in the single-angle image set to generate a fused image for damage recognition.

[0069] In this embodiment, step S5 combines an image sequence under multi-angle illumination with adaptively calculated weights by executing a deterministic weighted fusion algorithm to generate an optimized damage detection image. This solves the core technical problem of how to integrate scattered and of varying quality damage feature information under multiple illumination conditions into a high-quality image with complete features, uniform background, and best suitability for automatic recognition.

[0070] In some embodiments, the step of performing pixel-level fusion on all images in the single-angle image set based on each of the pixel-level weight maps to generate a fused image for damage identification specifically includes: S51. Obtain each single-angle image in the single-angle image set and the pixel-level weight map corresponding to each single-angle image; S52. For each pixel position, the pixel values ​​of all the single-angle images corresponding to that position are weighted and calculated with the weight value of that position in their respective pixel-level weight map, and the weighted results are summed to obtain the pixel value of that position in the fused image. S53. Repeat the weighted summation of step S52 for all pixel locations to generate the fused image.

[0071] In the embodiments of this application, the pixel-level fusion refers to an image synthesis method, which generates the final pixel value at each pixel position of the output image by calculating the weighted sum of the pixel values ​​of all input images at the corresponding position; the fused image is a single synthesized image finally generated by pixel-level fusion processing and used for subsequent damage identification analysis.

[0072] First, the system executes step S51 to acquire each single-angle image in the single-angle image set and the corresponding pixel-level weight map. This requires the system to accurately read two sets of data from the storage medium: one set is the original single-angle image set acquired and stored in step S2, denoted as {I_1,I_2,...,I_N}, where I_n represents the nth single-angle image; the other set is the pixel-level weight map set calculated and stored in step S4, denoted as {W_1,W_2,...,W_N}, where W_n is a weight map that corresponds one-to-one with I_n. The system ensures that the images and weight maps are perfectly aligned in order and spatial coordinates.

[0073] Next, the system enters the core calculation loop, performing a weighted calculation for each pixel location. This involves taking the pixel values ​​of all the single-angle images corresponding to that location and calculating the weight value of that location in their respective pixel-level weight maps. The weighted results are then summed to obtain the pixel value of that location in the fused image. For the first pixel location in the fused image, such as the top-left corner coordinate (0,0), the system performs the following operations: extracting the pixel value I_1(0,0) from the first single-angle image I_1 and the weight value W_1(0,0) from the corresponding weight map W_1, and calculating their product I_1(0,0)*W_1(0,0); then, performing the same operation on the second image and the weight map to obtain I_2(0,0)*W_2(0,0), and so on, until the calculation is completed for the Nth image; finally, all N weighted results are summed to obtain the pixel value Fused(0,0) at location (0,0) in the fused image.

[0074] For example, a point on a propeller blade may have a low pixel value (e.g., I_8=30) in the image of illumination unit 8 (side light) due to shadows caused by cracks, but the image has a high weight at this point (e.g., W_8=0.7); in the image of illumination unit 3 (front light), this point is brighter (e.g., I_3=200), but has a low weight (e.g., W_3=0.1). Therefore, the value of this point in the fused image is 30*0.7+200*0.1+...=21+20+..., and the dark crack features in the side light image are thus significantly preserved and enhanced.

[0075] Then, the system performs a weighted summation of step S52 on all pixel positions to generate the fused image. Following the raster scan order, the system repeatedly performs the weighted summation calculation of step S52 for each pixel position (x, y) in the image (where x ranges from 0 to image width - 1, and y ranges from 0 to image height - 1). After each position is calculated, the result Fused(x, y) is written to the corresponding position in the output image buffer. Once all pixel positions have been calculated, the complete image data stored in the output buffer is the final fused image.

[0076] This fused image effectively suppresses overexposed highlights or deep shadows that exist in some single-angle images due to complex curved surfaces and fixed light sources. At the same time, it gathers and enhances various damage features (such as cracks with different orientations and pits in different locations) that are scattered under different lighting angles into the same image. The result is an image with a uniform background and globally enhanced contrast of damage features, providing a near-ideal input for subsequent automated recognition algorithms.

[0077] S6. Analyze the fused image to identify surface damage to the propeller blades.

[0078] In this embodiment, step S6 performs automated damage localization, verification, and classification on the optimized fused image, thereby transforming the high-quality image information generated in the preceding steps into a structured inspection report that can directly guide maintenance. This solves the core technical problems of traditional visual inspection methods, which rely on manual interpretation, are inefficient, and have poor consistency, and achieves intelligent and objective identification of propeller blade surface damage.

[0079] In some embodiments, analyzing the fused image to identify surface damage to the propeller blades specifically includes: The fused image is subjected to damage feature enhancement and segmentation processing to obtain a binary mask image containing candidate damage regions; Based on predefined damage morphology and grayscale features, each candidate damage region in the binary mask image is verified and classified to determine whether the candidate damage region is a real damage and the damage type. The location and type information of the verified and classified damage areas are integrated and output to form the final damage identification result.

[0080] In some embodiments, the step of verifying and classifying each candidate damage region in the binary mask image based on predefined damage morphology and grayscale features, and determining whether the candidate damage region is a real damage and its damage type, specifically includes: Extract the morphological feature parameters and grayscale statistical feature parameters of the corresponding region in the fused image for each candidate damage region in the binary mask image; The morphological feature parameters and grayscale statistical feature parameters of the candidate damage region are compared with a preset damage feature threshold range. If they all fall within the corresponding preset damage feature threshold range, they are determined to be real damage. For areas determined to be real damage, they are further classified as crack damage or corrosion damage based on the morphological characteristic parameters.

[0081] In this embodiment, the binary mask image is an image containing only black and white pixel values, where the connected regions formed by white pixels represent potential damage regions segmented from the fused image; the candidate damage region refers to each independent connected region of white pixels in the binary mask image, which is considered a suspected damage target that needs further verification; the damage morphology and grayscale features are a set of parameters used to quantify the geometric shape and brightness distribution characteristics of a region; the final damage identification result is a structured data report, which includes the spatial location, category, and possible size information of each confirmed damage.

[0082] First, the system performs damage feature enhancement and segmentation processing on the fused image. Since the fused image generated in step S5 has significantly enhanced damage features, the system can employ a relatively stable global threshold segmentation algorithm. For example, using Otsu's method, an optimal threshold T_otsu is automatically calculated. The system sets pixels with grayscale values ​​below T_otsu in the fused image to white (foreground) and pixels with grayscale values ​​above or equal to T_otsu to black (background), thus initially obtaining a binary mask image that highlights dark damage regions. To further eliminate minor noise, morphological opening operations can be performed on the binary image, ultimately obtaining a clear binary mask image containing only significant candidate damage regions.

[0083] Next, the system performs key verification and classification of each candidate damage region in the binary mask image based on predefined damage morphology and grayscale features.

[0084] For each candidate damage region in the mask image, the system performs the following operations: extracting the morphological feature parameters and grayscale statistical feature parameters of the corresponding region of each candidate damage region in the binary mask image in the fused image.

[0085] Specifically, the system extracts corresponding pixel blocks on the original fused image based on the contour of the candidate region. The calculated morphological feature parameters include the region area A (number of pixels), the length L and width W of the minimum bounding rectangle, the aspect ratio R (R=L / W, L≥W), the perimeter P, and the roundness C (C=4πA / P²). The calculated gray-scale statistical feature parameters include the average gray value μ_region of the pixels in the region, and the contrast between the average gray value of the region and the average gray value μ_background of its immediate surrounding background region, Contrast. The calculation formula can be defined as Contrast=(μ_background-μ_region) / μ_background.

[0086] Then, the system compares the morphological feature parameters and grayscale statistical feature parameters of the candidate damage region with a preset damage feature threshold range. The system presets the threshold range that the real damage should meet. For example, the area A must be greater than the minimum damage area threshold A_min; the contrast Contrast must be greater than the minimum visibility threshold C_min. If both fall within the corresponding preset damage feature threshold range, it is determined to be a real damage.

[0087] For example, if a region satisfies A>A_min and Contrast>C_min, it passes the verification and is determined to be a real surface damage; otherwise, it is considered a false feature (such as residual water stains or image noise) and is filtered out. For real damage regions that pass the verification, the system performs further classification as crack damage or corrosion damage based on the morphological feature parameters.

[0088] The system's default classification rule is that if the aspect ratio R of a region is greater than a large threshold R_crack (e.g., R_crack=5) and its width W is less than the maximum crack width W_max, then the region is classified as crack damage; if the area A of a region is large, but the aspect ratio R is small (the shape is closer to a circle or an irregular block), and the roundness C is less than a threshold, then the region is classified as corrosion damage.

[0089] Finally, the system integrates and outputs the location and type information of the verified and classified damage areas. The system collects information on all verified and classified damage areas, including the coordinates of the vertices of their contour polygons, damage type labels, calculated area or length, etc., and encapsulates them into a structured JSON data file or directly visualizes and overlays them on the original image in the form of annotation boxes and labels to generate the final damage identification result.

[0090] In this embodiment, this step avoids the subjectivity and fatigue of manual inspection. It objectively verifies the candidate region through quantified feature thresholds, effectively distinguishes between real damage and image artifacts, greatly improves the accuracy and reliability of the detection results, and directly outputs typified and localized results, which greatly improves the efficiency of inspection.

[0091] like Figure 2 The diagram shown is a functional block diagram of a machine vision-based intelligent inspection method system for propeller blade surface damage provided in an embodiment of this application.

[0092] The intelligent inspection method system 100 for propeller blade surface damage based on machine vision described in this application can be installed in an electronic device. Depending on the functions implemented, the intelligent inspection method system 100 for propeller blade surface damage based on machine vision may include a coaxial illumination and image acquisition module 101, a model-driven illumination optimization acquisition module 102, a damage feature extraction and classification module 103, an adaptive weight map calculation module 104, a pixel-level image fusion module 105, and a damage intelligent recognition module 106. The modules described in this application can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0093] In this embodiment, the functions of each module / unit are as follows: The coaxial illumination and image acquisition module 101 is used to coaxially set the ring LED light source array with the image acquisition device and align it with the area to be inspected on the propeller blades. The model-driven lighting optimization acquisition module 102 is used to sequentially illuminate each lighting unit in the ring LED light source array according to the estimated three-dimensional curved surface model of the area to be inspected, and to simultaneously acquire a single-angle image of the area to be inspected by the image acquisition device each time it is illuminated, thereby obtaining a set of single-angle images. The damage feature extraction and classification module 103 is used to extract damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image. The adaptive weight map calculation module 104 is used to calculate the pixel-level weight map corresponding to each image in the single-angle image set based on each of the damage feature maps. The pixel-level image fusion module 105 is used to perform pixel-level fusion on all images in the single-angle image set according to the pixel-level weight maps to generate a fused image for damage recognition. The damage intelligent recognition module 106 is used to analyze the fused image to identify surface damage to the propeller blades.

[0094] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0095] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0096] Furthermore, the functional modules in the various embodiments of this application 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 in the form of hardware plus software functional modules.

[0097] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application.

[0098] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application.

Claims

1. A machine vision-based intelligent inspection method for surface damage on propeller blades, characterized in that, The method includes: S1. Set the ring-shaped LED light source array coaxially with the image acquisition device and align it with the area to be inspected on the propeller blades; S2. Based on the estimated three-dimensional surface model of the area to be inspected, each illumination unit in the ring LED light source array is lit sequentially in the optimized order, and each time it is lit, the image acquisition device simultaneously acquires a single-angle image of the area to be inspected, thereby obtaining a set of single-angle images. S3. Extract damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image; S4. Based on each of the damage feature maps, calculate the pixel-level weight map corresponding to each image in the single-angle image set; S5. Based on the pixel-level weight maps, perform pixel-level fusion on all images in the single-angle image set to generate a fused image for damage recognition; S6. Analyze the fused image to identify surface damage to the propeller blades.

2. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, The step of coaxially aligning the ring-shaped LED light source array with the image acquisition device and aiming it at the area to be inspected on the propeller blades specifically includes: The ring-shaped LED light source array is divided into multiple independent lighting sectors; Control multiple lighting sectors so that the lighting sectors are individually lit sequentially along a circular path around the optical axis of the image acquisition device; When each of the lighting sectors is illuminated, the image acquisition device is controlled to acquire a single-angle image of the area to be inspected. The single-angle images corresponding to all the lighting sectors are collected to form the single-angle image set.

3. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, Based on the estimated three-dimensional surface model of the area to be inspected, the illumination units in the annular LED light source array are sequentially lit in an optimized order. Each time the illumination is performed, the image acquisition device simultaneously acquires a single-angle image of the area to be inspected, thereby obtaining a set of single-angle images. Specifically, this includes: Based on the estimated three-dimensional surface model of the area to be inspected, the relationship between the illumination direction of each lighting unit in the ring LED light source array and the local normal vector of each point on the estimated three-dimensional surface model is analyzed to determine the lighting order of the lighting units. The lighting order of the lighting units is configured such that for any point in the area to be inspected, at least one lighting unit has an angle greater than 60 degrees between its illumination direction and the local normal vector of that point under the lighting order. Based on the lighting sequence of the lighting units, a lighting control command sequence corresponding one-to-one with each lighting unit in the ring LED light source array is generated; According to the lighting control command sequence, the corresponding lighting units are driven to emit light in sequence, and a synchronous trigger signal is sent to the image acquisition device at the same time; The image acquisition device responds to the synchronization trigger signal, acquires and outputs a single-angle image corresponding to the currently lit lighting unit; According to the driving order of the lighting units, the sequentially output single-angle images are sorted and cached to form a structured set of single-angle images.

4. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, The step of extracting damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image specifically includes: Each image in the single-angle image set is subjected to image enhancement processing to obtain an enhanced image; The enhanced image is subjected to edge detection processing to obtain the corresponding edge response map; The edge response map is subjected to binarization segmentation to obtain a preliminary segmented binary image; Morphological filtering is performed on the initially segmented binary image to eliminate isolated pixel regions with non-damaging features, resulting in a basic damage feature map. For each connected region in the basic damage feature map, the region is classified based on its morphological feature parameters, and the region is marked as a linear crack feature region or a planar corrosion feature region, thereby generating a damage feature map containing damage type information.

5. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 4, characterized in that, The step of marking the connected regions as linear crack feature regions or planar corrosion feature regions and generating a damage feature map containing damage type information specifically includes: Read the information of the pre-marked linear crack feature region and planar corrosion feature region in the damage feature map; A base weight gain coefficient is assigned to the pixels belonging to the linear crack feature region, which is greater than the base weight gain coefficient used by the pixels in the planar corrosion feature region.

6. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, The step of calculating the pixel-level weight map corresponding to each image in the single-angle image set based on each of the damage feature maps specifically includes: The damage feature maps are smoothed by a smoothing filter to obtain smoothed feature maps. Based on the smoothed feature values ​​of all images at the same pixel position in the single-angle image set, normalization calculation is performed on each smoothed feature map to obtain an initial weight map. The initial weight map is constrained to ensure that each initial weight value corresponding to each pixel position is non-negative and the sum is 1, thus obtaining the pixel-level weight map.

7. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, The step of performing pixel-level fusion on all images in the single-angle image set based on the pixel-level weight maps to generate a fused image for damage identification specifically includes: S51. Obtain each single-angle image in the single-angle image set and the pixel-level weight map corresponding to each single-angle image; S52. For each pixel position, the pixel values ​​of all the single-angle images corresponding to that position are weighted and calculated with the weight value of that position in their respective pixel-level weight map, and the weighted results are summed to obtain the pixel value of that position in the fused image. S53. Repeat the weighted summation of step S52 for all pixel locations to generate the fused image.

8. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 1, characterized in that, The analysis of the fused image to identify surface damage to the propeller blades specifically includes: The fused image is subjected to damage feature enhancement and segmentation processing to obtain a binary mask image containing candidate damage regions; Based on predefined damage morphology and grayscale features, each candidate damage region in the binary mask image is verified and classified to determine whether the candidate damage region is a real damage and the damage type. The location and type information of the verified and classified damage areas are integrated and output to form the final damage identification result.

9. The intelligent inspection method for propeller blade surface damage based on machine vision as described in claim 8, characterized in that, The process of verifying and classifying each candidate damage region in the binary mask image based on predefined damage morphology and grayscale features, and determining whether the candidate damage region is a real damage and its damage type, specifically includes: Extract the morphological feature parameters and grayscale statistical feature parameters of the corresponding region in the fused image for each candidate damage region in the binary mask image; The morphological feature parameters and grayscale statistical feature parameters of the candidate damage region are compared with a preset damage feature threshold range. If they all fall within the corresponding preset damage feature threshold range, they are determined to be real damage. For areas determined to be real damage, they are further classified as crack damage or corrosion damage based on the morphological characteristic parameters.

10. A machine vision-based intelligent inspection method and system for propeller blade surface damage, used to implement the machine vision-based intelligent inspection method for propeller blade surface damage as described in any one of claims 1-9, characterized in that, The system includes: The coaxial illumination and image acquisition module is used to coaxially set the ring LED light source array with the image acquisition device and align it with the area to be inspected on the propeller blades. The model-driven lighting optimization acquisition module is used to sequentially illuminate each lighting unit in the ring LED light source array according to the estimated three-dimensional curved surface model of the area to be inspected, and to simultaneously acquire a single-angle image of the area to be inspected by the image acquisition device each time it is illuminated, thereby obtaining a set of single-angle images. The damage feature extraction and classification module is used to extract damage features from each image in the single-angle image set to obtain a damage feature map corresponding to each image. An adaptive weight map calculation module is used to calculate the pixel-level weight map corresponding to each image in the single-angle image set based on each of the damage feature maps. A pixel-level image fusion module is used to perform pixel-level fusion on all images in the single-angle image set according to the pixel-level weight maps to generate a fused image for damage recognition. The damage intelligent recognition module is used to analyze the fused image to identify surface damage to the propeller blades.

Images