A component surface defect detection method based on feature perception

Abstract

The application discloses a kind of component surface defect detection methods based on feature perception, comprising:1) the RGB image of the component to be detected is collected;2) the RGB image is mapped through defect space, and defect space mapping chart is obtained;3) the key point coordinates of RGB image are obtained;Defect space mapping chart is respectively input into regional suggestion network and backbone network, and ROI region and region of interest ROIs are obtained;Defect space mapping chart is superpixel fast segmentation, and multiple background regions are obtained;4) find the intersection background region of ROIs, and carry out regional feature fusion;5) compare ROIs with key point coordinates, obtain double-neighbor ROI region feature, compare double-neighbor ROI region feature, and the comparison result vector is used as the feature vector of ROIs;6) the feature vector is sent into fully connected layer, and classification and frame regression are carried out, to identify the class and position of detection object and mark.The application realizes high-precision detection of component surface weak defects in industrial detection scene.

Description

Technical Field

[0001] This invention relates to the technical field of component defect detection, and in particular to a feature-based method for detecting surface defects in components, which achieves high-speed automatic detection, reduces operating costs, and improves the detection accuracy of product defects on industrial component production lines. Background Technology

[0002] Industrial image defect detection is a technology that uses machine vision equipment to acquire images and determine whether defects exist in the acquired images. With the increasing demands for production quality in automated manufacturing, industrial image detection technology for weak defects has important and widespread applications in industrial automated manufacturing. This paper addresses the problem of weak defect industrial image detection, categorizing existing methods into traditional image processing-based methods and deep learning-based methods. It provides a detailed introduction to existing methods, summarizes their advantages and disadvantages, and offers future prospects.

[0003] Traditional image processing-based methods involve manually designing features and classifiers, which require relatively demanding imaging environments and high contrast between defective and non-defective regions. They also have low noise, poor adaptability, and often require changes to thresholds or redesign of algorithms when the imaging environment or defect type changes.

[0004] This system automatically learns features from large amounts of data using deep learning methods, possessing sufficient learning data and high-performance computing units. It exhibits relatively strong adaptability and can cope with certain changes in industrial inspection environments. However, it currently lacks specific network designs for weakly defective industrial images, resulting in unsatisfactory performance.

[0005] Therefore, the current methods for defect detection in industrial components have very low accuracy and cannot meet industrial needs. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a feature-based method for detecting surface defects in components. This method enables automatic defect detection of industrial components without the need for training specialized personnel, reducing the investment of human and material resources. Furthermore, it maintains stable detection accuracy and speed, achieving high efficiency.

[0007] To achieve the above objectives, the technical solution provided by this invention is: a method for detecting surface defects in components based on feature perception, comprising the following steps:

[0008] 1) Acquire RGB images of the components to be tested;

[0009] 2) The RGB image is passed through a defect spatial mapping network to obtain a defect spatial mapping map, wherein the defect spatial mapping network includes a channel attention network and a spatial attention network;

[0010] 3) Locate the key points in the RGB image and obtain their coordinates;

[0011] The defect spatial map is input into the region proposal network to obtain the ROI region. At the same time, the defect spatial map is input into the backbone network to obtain the feature map, and ROI pooling is performed to obtain the region of interest (ROIs).

[0012] The defect space map is segmented quickly using superpixels to obtain multiple background regions;

[0013] 4) Based on the obtained multiple background regions and ROIs, find the intersecting background regions through the ROIs and perform region feature fusion, that is, fuse the features of the intersecting background regions with the ROIs;

[0014] 5) Based on the obtained key point coordinates, compare the ROIs with the key point coordinates to obtain two ROI regions of the same size in the horizontal and vertical directions, which are called double nearest neighbor ROI region features. Compare the double nearest neighbor ROI region features and use the comparison result vector as the feature vector of the region of interest ROIs.

[0015] 6) Feed the feature vector into the fully connected layer for classification and bounding box regression to identify and label the category and location of the detected object.

[0016] Furthermore, in step 1), the object to be detected is placed on the image acquisition device, and the optical system reaches the computer through the camera and image acquisition card, and the RGB image of the object to be detected is obtained through computer rendering technology.

[0017] Further, in step 2), the RGB image is input into the defect spatial mapping network, and channel concact is performed to obtain channel attention vectors. The channel attention vectors are then fed into a fully connected FC layer in parallel through maxPool and AvgPool, and the following operations are performed: First, a Squeeze operation is performed, which compresses the spatial dimension, that is, each two-dimensional feature map becomes a real number, which is equivalent to a pooling operation with a global receptive field, while keeping the number of feature channels unchanged; then, an Excitation operation is performed, which generates weights for each feature channel through the parameter channel attention module. The channel attention module is learned to explicitly model the correlation between feature channels. The channel attention module uses a 2-layer bottleneck structure, which is implemented by a fully connected layer that first reduces the dimension and then increases the dimension, plus a sigmoid function; after obtaining the weights of each feature channel, the corresponding weights are applied to each of the original feature channels. Based on the specific task, the importance of different channels is learned, and the defect channel attention map F is obtained.

[0018] Next, the defect channel attention map F is input into the spatial attention network. The defect channel attention map F output by the channel attention module is used as the input feature map of the spatial attention network, and then the following operations are performed: First, the feature map is subjected to max pooling (MaxPool) and average pooling (AvgPool) to obtain two weight vectors of H×W×1, where H is the height and W is the width. The number of channels in the feature map is changed from [C,H,W] to [1,H,W], where C is the feature layer. All channels of the same feature point are pooled to obtain two feature maps. Then, these two feature maps are concatenated based on the channel to form the feature map spatial weights of [2,H,W]. The feature map spatial weights are then subjected to a 7×7 convolution (7×7 is better than 3×3) to reduce the dimensionality to 1 channel, resulting in a feature map of [1,H,W]. This [1,H,W] feature map represents the importance of each point on the feature map, with larger values ​​being more important. The [1,H,W] feature map is then subjected to sigmoid to generate spatial attention. Finally, the spatial attention feature is multiplied by the input of the spatial attention network to obtain the generated map s. The map s is then superscaled into a defect spatial map.

[0019] Furthermore, in step 3), based on image features, contours are extracted using edge detection and prior image operations, followed by grayscale segmentation and then key point localization.

[0020] Furthermore, in step 3), the region proposal network is a deep network, consisting of convolutional layers, pooling layers, and nonlinear mapping layers;

[0021] The convolution process is as follows:

[0022]

[0023] In the formula, f1[x,y] represents the image data in the coordinate (x,y) region, w[x,y] is the convolution kernel, f2[x,y] is the feature obtained after convolving the image in the coordinate (x,y) region, and n i n j f[x+n] represents the offset distance from the center of the convolution, where n1 and n2 are the maximum offset distances in the vertical and horizontal directions, respectively. i ,y+n j ] is the image in (x+n i ,y+n j The value of w[n) i ,n j] is the convolution kernel in (n i ,n j The weight of the position;

[0024] Its nonlinear mapping process:

[0025] f3[x,y]=max(0,f2[x,y])

[0026] In the formula, f3[x,y] is the feature map obtained after performing nonlinear mapping;

[0027] In the RGB image input region proposal network, each point in the feature map of the RGB image is defined as an anchor point. Each anchor point defines 9 anchor boxes centered on itself. After removing the anchor boxes that exceed the image region, binary classification and bounding box regression are performed on the remaining anchor boxes.

[0028] a. Binary classification: y = f[f4(x,y)]

[0029] In the formula, y is the classification prediction of the foreground bounding box, f4(x,y) is the anchor box feature map, f is the classifier, the classifier is manually set with a threshold, the prediction of the foreground is greater than this threshold and is added to the subsequent calculation, the prediction of the background is less than this threshold and is discarded.

[0030] b. Boundary regression: r = [Δx, Δy, Δh, Δw] = g(f4[x, y])

[0031] In the formula, r is the offset of the foreground bounding box, g is the linear regression function; Δx and Δy are the predicted center offset of the anchor box; Δh and Δw are the anchor box scaling factors; the anchor box is adjusted in position and scale according to the foreground regression; then non-maximum suppression is used to filter the anchor boxes and remove overlapping anchor boxes; then the top n anchor boxes with the highest confidence are selected as the ROI region and proceed to the subsequent processing steps.

[0032] Furthermore, in step 3), the superpixel segmentation algorithm SLIC is used for linear iterative clustering. The color image is converted into a LAB color space vector through linear iterative clustering. Then, a new 5-dimensional feature vector is constructed from the LAB color space vector and the position vector of the image itself. Then, clustering is performed according to the distance standard constructed for the 5-dimensional feature vector. The image is segmented into multiple local background regions, i.e., multiple background regions, through clustering.

[0033] Furthermore, in step 4), the intersecting background regions are found using the coordinates of the ROIs, and the feature vectors of the background regions and the feature vectors of the ROI regions are fused to perform regional feature fusion.

[0034] Further, in step 5), the ROIs are compared with the previously calculated key point coordinates. Based on the range of the key point, two nearest neighbor ROI regions are found in the direction of the key point to obtain the features of the two nearest neighbor ROI regions. The features of the two nearest neighbor ROI regions are compared by difference, and the comparison result vector is used as the feature vector of the region of interest ROIs.

[0035] Furthermore, in step 6), the two tasks of classification and regression are completed:

[0036] Category: y′=max(h(f) p ))

[0037] In the formula, f p y′ represents the weighted fusion feature vector, h is the multi-classifier, and the output y′ is the confidence score of each class.

[0038] The feature vectors are classified, and the classification result with the highest confidence is taken as the classification result of the ROI region;

[0039] Regression: r′=[Δx′,Δy′,Δh′,Δw′]=g(f p )

[0040] In the formula, r′ is the offset of the predicted bounding box; Δx′, Δy′ are the predicted center offsets of the predicted bounding box; Δh′, Δw′ are the scaling factors of the predicted bounding box; and g is the linear regression function.

[0041] Regression is performed on each ROI region to obtain a more precise ROI region, that is, a more accurate category and location.

[0042] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0043] 1. Compared with other deep learning detection methods, this invention improves detection accuracy while maintaining detection speed. The proposed defect space mapping network maps the original image space to a defect image space with better defect representation, thereby enhancing the saliency of defect features. In particular, it is designed for weak defects, enabling the network to better identify weak defects and making the feature extraction and representation of the subsequent perception network more accurate.

[0044] 2. The Region Proposal Network of this invention enables the fusion of local background region information into the Region of Interest (ROI), providing better feature discrimination and interpretation for different features at different locations of defects. This avoids the blind identification problem in the application of traditional object detection in the field of defect detection, i.e., the need to distinguish defects in different parts of the same ROI.

[0045] 3. The dual-nearest neighbor ROI region feature designed in this invention homogenizes the adaptive features of defect features and enhances feature contrast. This solves the problem of weak differences in defect features, especially weak defect features, greatly improving the network's ability to detect weak defects. Furthermore, the network's generalization ability is improved, allowing it to generalize to image recognition under different lighting conditions. The homogenization of adaptive features enhances the mapping ability between the feature space and the solution space, resulting in higher detection accuracy for the method proposed in this invention.

[0046] 4. The method of the present invention has a wide range of applications in computer vision tasks, can realize end-to-end training and detection, has strong data adaptability, and has broad application prospects. Attached Figure Description

[0047] Figure 1 This is a design framework diagram of the method of the present invention.

[0048] Figure 2 This is a structural diagram of the defect space mapping network. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0050] like Figure 1 As shown in the figure, this embodiment discloses a method for detecting surface defects of components based on feature perception, the specific details of which are as follows:

[0051] 1) Place the object to be detected on the image acquisition device. The optical system reaches the computer through the camera and image acquisition card. The RGB image of the object to be detected is obtained through computer rendering technology.

[0052] 2) Input the RGB image into the defect spatial mapping network (the defect spatial mapping network includes a channel attention network and a spatial attention network; see the structure below). Figure 2As shown in the diagram, channel concact is performed to obtain channel attention vectors. These channel attention vectors are then fed into a fully connected FC layer in parallel through maxPool and AvgPool, and the following operations are performed: First, a Squeeze operation is performed, which compresses the spatial dimension, turning each two-dimensional feature map into a real number, equivalent to a pooling operation with a global receptive field, while keeping the number of feature channels unchanged. Next, an Excitation operation is performed, which generates weights for each feature channel through the parameter channel attention module. The channel attention module is learned to explicitly model the correlation between feature channels. The channel attention module uses a two-layer bottleneck structure, implemented by a fully connected layer that first reduces and then increases the dimension, plus a sigmoid function. After obtaining the weights for each feature channel, the corresponding weights are applied to each of the original feature channels. Based on a specific task, the importance of different channels is learned, resulting in the defective channel attention map F.

[0053] Next, the defect channel attention map F is input into the spatial attention network. The defect channel attention map F output by the channel attention module is used as the input feature map of the spatial attention network, and then the following operations are performed: First, the feature map is subjected to max pooling (MaxPool) and average pooling (AvgPool) to obtain two weight vectors of H×W×1, where H is the height and W is the width. The number of channels in the feature map is changed from [C,H,W] to [1,H,W], where C is the feature layer. All channels of the same feature point are pooled to obtain two feature maps. Then, these two feature maps are concatenated based on the channel to form the feature map spatial weights of [2,H,W]. The feature map spatial weights are then subjected to a 7×7 convolution (7×7 is better than 3×3) to reduce the dimensionality to 1 channel, resulting in a feature map of [1,H,W]. This [1,H,W] feature map represents the importance of each point on the feature map, with larger values ​​being more important. The [1,H,W] feature map is then subjected to sigmoid to generate spatial attention. Finally, the spatial attention feature is multiplied by the input of the spatial attention network to obtain the generated map s. The map s is then superscaled into a defect spatial map.

[0054] 3) Based on image features, edge detection and prior image operations are used to extract contours, then grayscale segmentation is performed, followed by key point localization.

[0055] The defect spatial map is input into the region proposal network to obtain the ROI region. At the same time, the defect spatial map is input into the backbone network to obtain the feature map, and ROI pooling is performed to obtain the region of interest (ROIs).

[0056] The region proposal network is a deep network, consisting of convolutional layers, pooling layers, and nonlinear mapping layers.

[0057] The convolution process is as follows:

[0058]

[0059] In the formula, f1[x,y] represents the image data in the coordinate (x,y) region, w[x,y] is the convolution kernel, f2[x,y] is the feature obtained after convolving the image in the coordinate (x,y) region, and n i n j f[x+n] represents the offset distance from the center of the convolution, where n1 and n2 are the maximum offset distances in the vertical and horizontal directions, respectively. i ,y+n j ] is the image in (x+n i ,y+n j The value of w[n) i ,n j ] is the convolution kernel in (n i ,n j The weight of the position;

[0060] Its nonlinear mapping process:

[0061] f3[x,y]=max(0,f2[x,y])

[0062] In the formula, f3[x,y] is the feature map obtained after performing nonlinear mapping;

[0063] In the RGB image input region proposal network, each point in the feature map of the RGB image is defined as an anchor point. Each anchor point defines 9 anchor boxes centered on itself. After removing the anchor boxes that exceed the image region, binary classification and bounding box regression are performed on the remaining anchor boxes.

[0064] a. Binary classification: y = f[f4(x,y)]

[0065] In the formula, y is the classification prediction of the foreground bounding box, f4(x,y) is the anchor box feature map, f is the classifier, the classifier is manually set with a threshold, the prediction of the foreground is greater than this threshold and is added to the subsequent calculation, the prediction of the background is less than this threshold and is discarded.

[0066] b. Boundary regression: r = [Δx, Δy, Δh, Δw] = g(f4[x, y])

[0067] In the formula, r is the offset of the foreground bounding box, g is the linear regression function; Δx and Δy are the predicted center offsets of the anchor boxes; Δh and Δw are the anchor box scaling factors; the anchor boxes are adjusted in position and scale according to the foreground regression; then, non-maximum suppression is used to filter the anchor boxes and remove overlapping anchor boxes; then, the top n anchor boxes with the highest confidence are selected as ROI regions and enter the subsequent processing steps; 2000 adaptive candidates are obtained, and ROI pooling is performed with the feature map to obtain ROIs.

[0068] The superpixel segmentation algorithm SLIC is used for linear iterative clustering. The color image is converted into a LAB color space vector through linear iterative clustering. Then, a new 5-dimensional feature vector is constructed from the LAB color space vector and the position vector of the image itself. Then, clustering is performed according to the distance criterion constructed for the 5-dimensional feature vector. The image is segmented into multiple local background regions, i.e., multiple background regions, through clustering.

[0069] 4) Find the intersecting background regions of ROIs using their coordinates, and fuse the feature vectors of the background regions with the feature vectors of the ROI regions.

[0070] 5) Compare the ROIs with the previously calculated keypoint coordinates. Based on the range of the keypoints, find two nearest neighbor ROI regions in the direction of the keypoints to obtain the features of the two nearest neighbor ROI regions. Perform a difference comparison on the features of the two nearest neighbor ROI regions and use the comparison result vector as the feature vector of the region of interest ROIs.

[0071] 6) Feed the feature vectors into the fully connected layer for classification and bounding box regression to identify and label the category and location of the detected objects; the classification and regression tasks are completed as follows:

[0072] Category: y′=max(h(f) p ))

[0073] In the formula, f p y′ represents the weighted fusion feature vector, h is the multi-classifier, and the output y′ is the confidence score of each class.

[0074] The feature vectors are classified, and the classification result with the highest confidence is taken as the classification result of the ROI region;

[0075] Regression: r′=[Δx′,Δy′,Δh′,Δw′]=g(f p )

[0076] In the formula, r′ is the offset of the predicted bounding box; Δx′, Δy′ are the predicted center offsets of the predicted bounding box; Δh′, Δw′ are the scaling factors of the predicted bounding box; and g is the linear regression function.

[0077] Regression is performed on each ROI region to obtain a more precise ROI region, that is, a more accurate category and location.

[0078] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A feature-aware based surface defect detection method for a component, characterized in that, The method comprises the following steps: 1) collecting an RGB image of a component to be detected; 2) mapping the RGB image through a defect space mapping network to obtain a defect space mapping image, wherein the defect space mapping network comprises a channel attention network and a spatial attention network; 3) positioning key points of the RGB image to obtain key point coordinates; inputting the defect space mapping image into a region proposal network to obtain an ROI region, and inputting the defect space mapping image into a backbone network to obtain a feature map, and performing ROI pooling processing to obtain ROIs; performing superpixel fast segmentation on the defect space mapping image to obtain multiple background regions; 4) according to the obtained multiple background regions and ROIs, finding the intersection background regions of the ROIs, and performing region feature fusion, that is, fusing the intersection background regions and ROI features; 5) according to the obtained key point coordinates, comparing the ROIs and the key point coordinates, obtaining two ROI regions with the same size in the horizontal and vertical directions, respectively, called double-neighbor ROI region features, comparing the double-neighbor ROI region features, and taking the comparison result vector as a feature vector of the ROIs; 6) inputting the feature vector into a fully connected layer to perform classification and bounding box regression, and identifying the class and position of the detection object and marking.

2. The feature-aware based surface defect detection method of claim 1, wherein, In step 1), the detection object is placed on an image acquisition device, an optical system reaches a computer through a camera and an image acquisition card, and an RGB image of the detection object is obtained through computer rendering technology.

3. The feature-aware based surface defect detection method of claim 2, wherein, In step 2), the RGB image is input into the defect space mapping network and concatted to obtain a channel attention vector, which is input into a fully connected FC layer in parallel after being processed by maxPool and AvgPool, and then the following operations are performed: first, a squeeze operation is performed, which compresses the spatial dimension, that is, each two-dimensional feature map becomes a real number, which is equivalent to a pooling operation with a global receptive field, and the number of feature channels remains unchanged; then, an excitation operation is performed, which generates a weight for each feature channel through a parameter channel attention module, the channel attention module is learned to explicitly model the correlation between feature channels, and the channel attention module uses a 2-layer bottleneck structure, a full connection layer for reducing dimension and then increasing dimension + a Sigmoid function to realize; after obtaining the weight of each feature channel, the corresponding weight is applied to each original feature channel, the importance of different channels is learned based on a specific task, and a defect channel attention mapping image F is obtained; Then, the defect channel attention map F is input into the spatial attention network, and the defect channel attention map F output by the channel attention module is taken as the input feature map of the spatial attention network, and then the following operations are performed: first, the input feature map is subjected to maximum pooling MaxPool and average pooling AvgPool to obtain two HxWx1 weight vectors, H is the height, and W is the width; the channel number of the input feature map changes from [C, H, W] to [1, H, W], C is the feature map layer, and all channels of the same feature point are pooled to obtain two feature maps; then the two feature maps are concatenated based on the channel to form a [2, H, W] feature map spatial weight, and then the feature map spatial weight is subjected to a 7x7 convolution operation to reduce the dimension to 1 channel to obtain a [1, H, W] feature map, which represents the importance of each point on the feature map, and the larger the value, the more important, and the [1, H, W] feature map is subjected to sigmoid to generate a spatial attention feature, and finally the spatial attention feature and the input of the spatial attention network are multiplied to obtain a generated mapping s, which is super-scaled into a defect spatial mapping.

4. The method according to claim 3, wherein: In step 3), according to the image features, the edge detection and prior image operation are used to extract the contour, and then the gray scale segmentation is performed, and then the key point positioning is performed.

5. The method of claim 4, wherein: In step 3), the region proposal network is a deep network composed of convolutional layers, pooling layers and nonlinear mapping layers; The convolution process is as follows: ； In the formula, This represents the data of the image in the coordinate (x, y) region. For convolution kernel, The features obtained after convolving the image in the (x,y) coordinate region are... , This is the offset distance from the center of the convolution. , These represent the maximum offset distance in the vertical direction and the maximum offset distance in the horizontal direction, respectively. For the image in ( The value of ) For the convolution kernel in ( The weight of the position; The nonlinear mapping process is as follows: ； In the formula, is the feature map obtained after the nonlinear mapping; The RGB image is input into the region proposal network, and each point of the feature map in the RGB image is defined as an anchor point, and each anchor point defines 9 anchor boxes with itself as the center. Except for the anchor boxes that exceed the image area, the remaining anchor boxes are subjected to binary classification and bounding box regression. a. Binary classification: ; In the formula, is a classification prediction of the foreground bounding box, is an anchor feature map, is a classifier, the classifier is set a threshold, the prediction greater than the threshold is foreground and is added to the subsequent step calculation, the prediction less than the threshold is background and is discarded; b. bounding box regression: ; In the formula, is an offset of the foreground frame, is a linear regression function; is a center offset prediction of the anchor frame; is an anchor frame scale factor; the anchor frame is adjusted in position and scale according to the foreground regression; then the anchor frame is screened using non-maximum suppression to remove overlapping anchor frames; and the top n anchor frames with the highest confidence are taken as the ROI region to enter the subsequent step for processing.

6. The method according to claim 5, wherein: In step 3), the linear iterative clustering algorithm SLIC is used to perform linear iterative clustering, the color image is converted into a LAB color space vector through linear iterative clustering, then a new 5-dimensional feature vector is constructed from the LAB color space vector and the position vector of the image itself, and then the image is segmented into multiple local background regions, i.e. multiple background regions, according to the distance criterion constructed for the 5-dimensional feature vector.

7. The method according to claim 6, wherein: In step 4), the intersection background region is found through the ROI coordinates, and the feature vector of the background region is fused with the feature vector of the ROI region.

8. The feature-aware based surface defect detection method of components according to claim 7, characterized in that: In step 5), the ROIs are compared with the key point coordinates calculated before, and according to the range of the key point, two adjacent ROI regions are found in the direction of the key point to obtain double-adjacent ROI region features, and the double-adjacent ROI region features are compared to obtain a feature vector of the region of interest ROI.

9. The feature-aware based surface defect detection method of components according to claim 8, characterized in that: In step 6), two tasks of classification and regression are completed: Classification: ; wherein is the weighted fused feature vector, h is the multi-classifier, and output is the confidence for each class. The feature vectors are classified, and the classification result with the highest confidence is taken as the classification result of the ROI region; Regression: ; In the formula, is a prediction of the offset of the bounding box; is a prediction of the center offset prediction of the bounding box; is a prediction of the scale factor of the bounding box; g is a linear regression function; Each ROI region is regressed to obtain a more accurate ROI region, i.e., a more accurate category and position.

Images