Multi-scale object detection method guided by image pyramid features

Abstract

The present invention discloses a multi-scale target detection method guided by image pyramid features, comprising the following steps: S1, using a color image as network input, FPN as the target detection framework, and employing a sorted downsampling method to extract image features; S2, using the same color image as input, employing a constructed dual-bottleneck sub-convolutional network to extract position information and detail features at each level in the image pyramid; S3, inputting the image features of each level extracted in step S2 and the corresponding deep features of the backbone network into a constructed hierarchical feature fusion module to fuse high-resolution, weak semantic features with low-resolution, strong semantic features; and S4, introducing a focal loss reconstruction function to complete target detection. This method not only enhances spatial position information but also avoids the loss of a large amount of detail information during downsampling, thereby improving the target detection network's ability to discern small and nearby targets.

Description

Technical Field

[0001] The present invention relates to a multi-scale target detection method, and in particular to a multi-scale target detection method guided by image pyramid features. Background Art

[0002] The task of object detection is to accurately predict the category and coordinate location of each object in a natural image. It is widely used in many fields, from autonomous driving to smart communities to video surveillance, and holds great research value. However, objects of different categories can have similar appearances and sizes, while objects of the same category can also differ significantly in appearance and size. Complex and diverse backgrounds, coupled with the frequent occlusion of objects, make object detection one of the most challenging tasks in computer vision. Traditional object detection methods extract features from images using manually designed feature extraction operators. However, these features have low representational power and poor generalization, which has limited their further development.

[0003] In recent years, the emergence of deep learning has greatly promoted the development of computer vision. Object detection algorithms based on convolutional neural networks (CNNs) extract target features from images through convolution operations. This enables the network to obtain more advantageous and deeper features and better handle complex situations such as occlusion, deformation, and illumination changes. Currently, object detection algorithms based on CNNs can be divided into two categories: two-step object detection algorithms based on region proposals and one-step object detection algorithms based on regression.

[0004] However, object detection algorithms still have many deficiencies, especially in multi-scale object detection. Existing methods cannot effectively detect the same object at different scales, especially small objects and nearby objects, whose semantic information disappears as the network deepens. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a multi-scale target detection method guided by image pyramid features that can improve the recognition of small targets and adjacent targets, and is used to solve the problem that high-level feature maps in multi-scale target detection easily confuse adjacent target features and ignore small targets.

[0006] Technical solution: The multi-scale target detection method of the present invention includes the following steps:

[0007] S1 uses color images as network input, uses FPN based on the ResNet-101 backbone network as the target detection framework, and adopts the sorting downsampling method to extract image features;

[0008] S2, taking the same color image in step S1 as input, uses the constructed double-bottleneck sub-convolutional network to extract the position information and detail features of each level in the image pyramid;

[0009] S3, inputting the image features of each layer extracted in step S2 and the deep features corresponding to the backbone network into the constructed hierarchical feature fusion module to complete the fusion of high-resolution, weak semantic features and low-resolution, strong semantic features;

[0010] S4 introduces Focal loss to reconstruct the loss function, trains multiple tasks, and completes target detection.

[0011] In step S1, the implementation process of the sorting downsampling method is as follows:

[0012] S11, slide a sliding window with a set step size on the feature map of the feature sampling layer of the convolutional neural network, sort the values ​​in the sliding window in ascending order, extract the four values ​​in the sliding window in sequence, and generate four new feature maps; the width and height of each new feature map are half of the original feature map, so the output of the sorted downsampling method is:

[0013]

[0014] in, Represents the feature map of each sampling layer of the convolutional neural network, W, H and D represent the width, height and number of channels of the feature map respectively, l is the level index of the convolutional neural network sampling layer; M j (·) represents the process of extracting the jth value in the sliding window, and four values ​​are extracted in each sliding window in sequence; represents the new feature map of the jth output in the lth downsampling layer, and each downsampling layer generates four new feature maps;

[0015] S12, the four new feature maps are juxtaposed and then input into a small convolutional network for feature refinement and channel adjustment; the final feature map output As the input of the next layer of the backbone network, W′, H′, and D′ represent the width, height, and number of channels of the final feature map, respectively.

[0016] In step S2, the process of constructing the double-bottleneck sub-convolutional network is as follows:

[0017] S21, define the input of the double bottleneck sub-convolutional network as:

[0018]

[0019] in, Indicates height H * , width is W *The image is also the input image of the target detection model; i is the level index of the image pyramid and the backbone network;

[0020] S22, input the i-th layer image in the image pyramid into the double-bottleneck sub-convolutional network, and extract the edge features of the image surface layer through a 5×5 convolution kernel and a 3×3 convolution kernel;

[0021] S23, the extracted edge features are input into a residual network unit with a two-bottleneck structure to extract detail features, and a side connection with a 1×1 convolution kernel is used to transmit the accurately located edge information to the extracted texture detail features;

[0022] The bottleneck structure consists of two 1×1 convolution kernels for feature map channel dimensionality reduction and dimensionality increase, and two 3×3 convolution kernels for learning shallow features;

[0023] S24, obtains a feature map with the same scale as the corresponding backbone network layer, which is the output of the residual network unit;

[0024] S25, taking images of different scales as input, defines the output of the double-bottleneck sub-convolutional network as:

[0025]

[0026]

[0027] in, Represents the features extracted from the image at layer i of the image pyramid; Represents the set of features extracted from all levels of images in the image pyramid.

[0028] In step S3, the hierarchical feature fusion module adopts a feature fusion module based on element-by-element addition; the output of element-by-element addition is defined as:

[0029]

[0030] in, and Represent two 3×3 convolution units, used as parameterized feature maps of the feature map, is a 1×1 convolution unit, which is used as a linear transformation of the feature map; BN[·] is a batch normalization operation of the convolution feature; T(·) represents a bilinear interpolation operation of the channel dimension, which is used to adjust the channel dimension of two different types of features; h(·) and g(·) are the output feature map of the double-bottleneck sub-convolution network and the feature map of the backbone network, respectively, and i is the hierarchical index of the image pyramid and the backbone network; I0 and I i Represent the original image and the image of the i-th layer in the image pyramid respectively.

[0031] In step S4, the classification loss function is as follows:

[0032]

[0033] Among them, p and p * are sample prediction value and sample true value respectively; α t ∈(0, 1) is a weight factor introduced for class 1, 1-α t is to introduce a weight factor for class-1; (1-p)γ is the modulation coefficient;

[0034] The position regression loss term is expressed as:

[0035]

[0036] Among them, L reg (t,t * ) is represented by a smooth L1 loss; t = {x, y, w, h} represents the bounding box position information of the sample prediction, where {x, y} represents the center coordinates of the bounding box, and {w, h} represents the width and height of the bounding box; t * is the sample label of t;

[0037] The target detection loss function is expressed as:

[0038]

[0039] in, is the number of samples for classification, is the number of samples for regression; w is the number of batches of training images; k is the index of a single sample in each batch of training samples; λ is the loss balance term.

[0040] Compared with the prior art, the present invention has the following significant effects:

[0041] 1. This paper uses a dual-bottleneck sub-convolutional network to extract shallow features at each layer of a multi-scale image pyramid. It also utilizes a hierarchical feature fusion module to introduce high-resolution shallow features rich in detail and position information into the target detection model. It also utilizes a novel sorting downsampling method to preserve a large amount of detail information that would otherwise be lost during the downsampling process. The reconstructed loss function alleviates the imbalance between foreground and background classification.

[0042] 2. The present invention makes full use of the shallow high-resolution features of the multi-scale image pyramid, while also avoiding the loss of a large amount of detail information during downsampling, thereby increasing the recognition of small and adjacent targets by the target detection network and improving model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1Schematic diagram of the double-bottleneck feature pyramid network (DBFP-Net) structure of the present invention;

[0044] Figure 2 Schematic diagram of the sorting downsampling method of the present invention;

[0045] Figure 3 Schematic diagram of the implementation of the sorting downsampling method of the present invention.

[0046] Figure 4 This is a double-bottleneck sub-convolutional network framework diagram of the present invention;

[0047] Figure 5 (a) is a schematic diagram of a feature fusion module based on element-by-element addition in the present invention.

[0048] (b) is a schematic diagram of a feature fusion module based on element-by-element multiplication in the present invention,

[0049] (c) is a schematic diagram of a feature fusion module based on feature map juxtaposition in the present invention;

[0050] Figure 6 Schematic diagram of some detection results of DBFP-Net on MS COCO. DETAILED DESCRIPTION

[0051] The present invention will be described in further detail below with reference to the accompanying drawings and specific implementations.

[0052] Figure 1 This is a schematic diagram of the structure of the Double Bottleneck Feature Pyramid Network (DBFP-Net). First, FPN (Feature Pyramid Networks) is used as the target detection framework, and ResNet-101 is used as the backbone network of FPN to extract deep features. A color image is input, and a new sorted downsampling feature enhancement method is used to replace the original downsampling method in the backbone network. This retains all values ​​within the sliding window during downsampling, alleviating the problem of deep features losing a large amount of detailed information. Then, a double-bottleneck sub-convolutional network is constructed to extract image features at each level in the image pyramid. A hierarchical feature fusion module is constructed to fuse the high-resolution, weak semantic features extracted by the double-bottleneck sub-convolutional network with the low-resolution, strong semantic features deep in the backbone network. This can provide the target detection network with shallow features extracted from the image pyramid and more spatial information. Finally, the Focalloss reconstruction loss function is introduced to alleviate the imbalance between foreground and background classification.

[0053] This paper uses FPN as the target detection framework and ResNet-101 as the backbone network to extract deep features. It inputs a color image and uses a new sorted downsampling feature enhancement method to replace the original downsampling method in the backbone network. It retains all the values ​​​​in the sliding window during downsampling and alleviates the problem of losing a large amount of detail information in deep features.

[0054] (1) Sorting downsampling method

[0055] In the continuous downsampling operation of CNN (convolutional neural network), the maximum pooling method is commonly used. The scale of the new feature map generated is smaller than the original feature map, which causes the new feature map to lose a lot of original feature map information. This is very disadvantageous for multi-scale target detection tasks that require a large amount of feature information to identify multiple types of targets. Therefore, the sorted downsampling feature enhancement method proposed in this invention is as follows: Figure 2 As shown in Figure 1, a 2×2 window with a stride of 2 is slid across the feature map of the CNN's feature sampling layer. The values ​​within the window are sorted in ascending order, and the four values ​​within the window are extracted sequentially. This preserves all the values ​​within the window. Finally, four new feature maps are generated, each with half the width and height of the original feature map, but without losing any information from the original feature map. The output of the sorted downsampling method is defined as:

[0056]

[0057] in, Represents the feature map of each sampling layer of the convolutional neural network (CNN), W, H and D represent the width, height and number of channels of the feature map respectively, l is the level index of the CNN sampling layer; M j (·) represents the process of extracting the jth value in the sliding window using the sorted downsampling method, where four values ​​are extracted in each sliding window in sequence; represents the new feature map of the j-th output in the l-th downsampling layer, and each downsampling layer generates four new feature maps.

[0058] The specific implementation of the sorting downsampling method is as follows Figure 3 The four new feature maps are juxtaposed and then input into a small convolutional network for feature refinement and channel adjustment. In order to keep consistent with the output of the original sampling layer of the backbone network, the small convolutional network consists of a 1×1 convolution kernel and a 3×3 convolution kernel. Output the final feature map As the input of the next layer of the backbone network, W′, H′, and D′ represent the width, height, and number of channels of the final feature map, respectively.

[0059] (2) Double Bottleneck Sub-Convolutional Network

[0060] A dual-bottleneck sub-convolutional network is constructed to extract image features at each level of the image pyramid, generating feature maps rich in spatial information at different scales. Because the dual-bottleneck sub-convolutional network is lightweight and shares parameters across all levels of the image pyramid, the additional computational complexity and parameter storage loss are negligible compared to the backbone network.

[0061] Figure 4 The following is a diagram of the dual-bottleneck sub-convolutional network. The input of the dual-bottleneck sub-convolutional network is a simple multi-scale image pyramid. An image pyramid is a set of images arranged in a pyramid shape with gradually decreasing resolution. They are obtained by sequentially downsampling the same input image. The input of the dual-bottleneck sub-convolutional network is defined as:

[0062]

[0063] in, Indicates resolution is H * ×W * (Height H * , width is W * ) image, which is also the input image of the backbone network; i is the hierarchical index of the image pyramid and the backbone network. First, the i-th layer image in the image pyramid is input into the double bottleneck sub-convolutional network, and the edge features of the image surface are extracted through a 5×5 convolution kernel and a 3×3 convolution kernel. Then, the extracted edge features are input into the residual network unit with two bottleneck structures to extract detail features, and the side connection with a 1×1 convolution kernel is used to transmit the accurately positioned edge information to the extracted texture detail features. The bottleneck structure consists of two 1×1 convolution kernels for feature map channel dimensionality reduction and dimensionality increase, and two 3×3 convolution kernels for learning shallow features. Finally, the output of the residual network unit is a feature map with the same scale as the corresponding backbone network level. Taking images of different scales as input, the output of the double bottleneck sub-convolutional network is defined as:

[0064]

[0065]

[0066] in, Represents the features extracted from the image at level i of the image pyramid. Represents the set of features extracted from all levels of images in the image pyramid.

[0067] (3) Hierarchical feature fusion module

[0068] A hierarchical feature fusion module is constructed to fuse the high-resolution, weak semantic features extracted by the double-bottleneck sub-convolutional network with the low-resolution, strong semantic features in the deep layer of the backbone network, and transmit them to the feature map of the corresponding layer of the FPN. This can provide the target detection network with shallow features extracted from the image pyramid and more spatial information. Three fusion methods are designed, such as Figure 5 As shown in (a), (b), and (c), their outputs are uniformly defined as:

[0069]

[0070] Among them, h(·) and g(·) are the output feature map of the double-bottleneck sub-convolutional network and the feature map of the backbone network respectively, i is the level index of the image pyramid and the backbone network, O i is the output feature of the feature fusion module on the i-th layer; I0 and I i Represent the original image and the image of the i-th layer in the image pyramid respectively. is the fusion equation of the feature fusion module. Generally, if the image pyramid has a total of (n-1) downsampled images, then the number of layers of the image pyramid is n.

[0071] Feature fusion modules based on element-by-element addition such as Figure 5 As shown in (a), the output feature h of the double bottleneck sub-convolutional network is i (I i ) and backbone network features g i (I0) is added. Since the dual bottleneck sub-convolutional network parameters are shared, it is necessary to adjust the channel dimensions of these two different types of features. The channel dimension bilinear interpolation operation T(·) is used to adjust the channel dimensions of these two different types of features. Therefore, the output O of element-by-element addition is defined as 1i for:

[0072]

[0073] in, and Represent two 3×3 convolution units, used as parameterized feature maps of the feature map, is a 1×1 convolution unit used for linear transformation of feature maps. BN[·] is the batch normalization operation of convolution features.

[0074] The feature fusion module based on element-by-element multiplication is as follows Figure 5 As shown in (b), the output feature h of the double bottleneck sub-convolutional network is i (I i ) and backbone network features g i(I0) multiplied. Then, it is refined by a 1×1 convolution kernel. Finally, through a side connection, the refined features are added to the corresponding backbone network features again. Define the output O of the element-by-element multiplication 2i for:

[0075]

[0076] in, 1×1 convolutional unit is used for feature refinement; The 1×1 convolution unit is used as a linear transformation of the feature map.

[0077] Feature fusion modules based on feature map juxtaposition Figure 5 As shown in (c), it is similar to the feature fusion method used in U-Net. The present invention also defines a feature map juxtaposition method, which outputs O 3i for:

[0078]

[0079] Among them, concat(·) represents the feature map concatenation operation.

[0080] The above methods demonstrate the flexibility of the hierarchical feature fusion module. Different fusion methods can be selected as part of the model of the present invention.

[0081] To compare the effectiveness of the three different fusion methods, the experiment implemented different feature fusion methods in the same object detection model. The comparison results are shown in Table 1.

[0082] Table 1 Comparison of hierarchical feature fusion methods

[0083]

[0084] The feature fusion method based on element-by-element addition achieved 44.7AP; the method based on element-by-element multiplication and feature map juxtaposition achieved 42.9AP and 43.6AP respectively. In small target detection, these three feature fusion methods achieved similar detection results. However, in medium-scale and large-scale target detection, the target detection model configured with the element-by-element addition fusion method improved by 4.1% and 3.6% respectively compared with the target detection model configured with the element-by-element multiplication fusion method. Table 1 shows that the feature fusion method using element-by-element addition has a better performance in AP. S 、AP M With AP L The evaluation criteria all achieved better detection results than the other two methods. Therefore, the present invention selects the feature fusion method based on element-by-element addition.

[0085] (4) Loss function

[0086] In order to alleviate the extreme imbalance between foreground and background, the Focal loss reconstruction loss function is introduced. The classification loss function is as follows:

[0087]

[0088] Among them, p and p * are the sample prediction value and sample true value respectively. t ∈(0, 1) is a weight factor introduced for class 1, 1-α t It is to introduce a weight factor for class-1 to adjust the weight of foreground and background classification. (1-p) γ is the modulating factor. This scaling factor can automatically reduce the weight of easy examples during training and quickly focus the model on difficult examples. In other words, the higher the confidence of the sample prediction, the smaller the contribution of the sample to the overall loss. As with Focal loss, we empirically set γ to 2 and α to 1. t Set to 0.25.

[0089] The position regression loss term is specifically expressed as:

[0090]

[0091] Among them, L reg (t,t * ) is represented by a smooth L1 loss; t = {x, y, w, h} represents the bounding box position information of the sample prediction, where {x, y} represents the center coordinates of the bounding box, and {w, h} represents the width and height of the bounding box; t * is the sample label of t.

[0092] The target detection loss function used in the present invention can be expressed as follows:

[0093]

[0094] in, is the number of samples for classification, is the number of regression samples. w is the number of batches of training images, k is the index of a single sample in each batch of training samples; λ is the loss balance term, which is empirically set to 2.

[0095] To further verify the performance of the proposed model, we compare the accuracy of DBFP-Net with existing object detection algorithms on the MS COCO dataset, as shown in Table 2.

[0096] Table 2 Comparison results with existing models on the COCO dataset

[0097]

[0098] As can be seen, the SSD model has the lowest AP value, at only 28.8%. SSD constructs a feature pyramid structure based on R-FCN, predicting objects at multiple levels without combining features or scores. This improves detection speed and accuracy, but the recall rate for small objects is still unsatisfactory. FPN uses an encoder-decoder network structure with lateral connections to associate low-level feature maps across resolutions and semantic levels, providing high-level semantic information for shallow network features, thereby improving the robustness of the detection model for multi-scale objects. The present invention introduces shallow features of a multi-scale image pyramid based on FPN to compensate for the information lost by the backbone network, exceeding FPN in AP by 8.5%. RetinaNet and the multi-scale object detection model guided by image pyramid features are both based on the FPN framework and use Focal Loss as the loss function, but DBFP-Net significantly improves RetinaNet by 5.6% (44.7AP vs. 39.1AP). Compared to the current best-performing detection model based on the feature pyramid method, M2Det, the present invention still exceeds it by 3.7% (44.7AP vs. 41.0AP). Compared to the recently popular anchor-free target detection methods, DBFP-Net surpasses CornerNet, FCOS, and FSAF by 4.2% (44.7AP vs. 40.5AP), 2.6% (44.7AP vs. 42.1AP), and 1.8% (44.7AP vs. 42.9AP) respectively. S 、AP M With AP L In terms of performance, the present invention is the best compared with other models in Table 2.

[0099] Figure 6 The detection results of the present invention on the MS COCO dataset are shown. The results show that for multi-scale targets in the dataset, the present invention retains more information for the objects and can accurately present their positions and classifications. DBFP-Net can not only detect extremely small targets, but also easily detect occluded and dense targets. For example, in the second picture in the third row and the fourth picture in the fifth row, the present invention detects the occluded bird very well; in the second picture in the second row and the third picture in the fifth row, the present invention can accurately detect dense fruits and people. The present invention is also highly robust to small targets with high-speed motion and pixel blur, such as the baseball in the second picture in the first row. In addition, some objects with sparse features can also be accurately detected, such as skateboards and cups.

[0100] In summary, the proposed method achieves an 8.5% improvement over the classic FPN network and can effectively detect occluded objects and objects of different sizes.

Claims

1. A multi-scale object detection method guided by image pyramid features, characterized in that: The steps are as follows: S1 uses color images as network input, uses FPN based on the ResNet-101 backbone network as the target detection framework, and adopts the sorting downsampling method to extract image features; S2, taking the same color image from step S1 as input, uses the constructed double-bottleneck sub-convolutional network to extract the position information and detail features of each level in the image pyramid; the process of constructing the double-bottleneck sub-convolutional network is as follows: S21, define the input of the double bottleneck sub-convolutional network as: in, Indicates height H * , width is W * The image is also the input image of the target detection model; i is the level index of the image pyramid and the backbone network; S22, input the i-th layer image in the image pyramid into the double-bottleneck sub-convolutional network, and extract the edge features of the image surface layer through a 5×5 convolution kernel and a 3×3 convolution kernel; S23, the extracted edge features are input into a residual network unit with a two-bottleneck structure to extract detail features, and a side connection with a 1×1 convolution kernel is used to transmit the accurately located edge information to the extracted texture detail features; The bottleneck structure consists of two 1×1 convolution kernels for feature map channel dimensionality reduction and dimensionality increase, and two 3×3 convolution kernels for learning shallow features; S24, obtains a feature map with the same scale as the corresponding backbone network layer, which is the output of the residual network unit; S25, taking images of different scales as input, defines the output of the double-bottleneck sub-convolutional network as: in, Represents the features extracted from the image at layer i of the image pyramid; Represents the set of features extracted from all levels of images in the image pyramid; S3, inputting the image features of each layer extracted in step S2 and the deep features corresponding to the backbone network into the constructed hierarchical feature fusion module to complete the fusion of high-resolution, weak semantic features and low-resolution, strong semantic features; S4 introduces Focal loss to reconstruct the loss function, trains multiple tasks, and completes target detection.

2. The multi-scale object detection method guided by image pyramid features according to claim 1, characterized in that: In step S1, the implementation process of the sorting downsampling method is as follows: S11, slide a sliding window with a set step size on the feature map of the feature sampling layer of the convolutional neural network, sort the values ​​in the sliding window in ascending order, extract the four values ​​in the sliding window in sequence, and generate four new feature maps; the width and height of each new feature map are half of the original feature map, so the output of the sorted downsampling method is: in, Represents the feature map of each sampling layer of the convolutional neural network, W, H and D represent the width, height and number of channels of the feature map respectively, l is the level index of the convolutional neural network sampling layer; M j (·) represents the process of extracting the jth value in the sliding window, and four values ​​are extracted in each sliding window in sequence; represents the new feature map of the jth output in the lth downsampling layer, and each downsampling layer generates four new feature maps; S12, the four new feature maps are juxtaposed and then input into a small convolutional network for feature refinement and channel adjustment; the final feature map output As the input of the next layer of the backbone network, W ′ 、H ′ and D ′ Represent the width, height and number of channels of the final feature map respectively.

3. The multi-scale object detection method guided by image pyramid features according to claim 1, characterized in that: In step S3, the hierarchical feature fusion module adopts a feature fusion module based on element-by-element addition; the output of element-by-element addition is defined as: in, and Represent two 3×3 convolution units, used as parameterized feature maps of the feature map, is a 1×1 convolution unit, which is used as a linear transformation of the feature map; BN[·] is a batch normalization operation of the convolution feature; T(·) represents a bilinear interpolation operation of the channel dimension, which is used to adjust the channel dimension of two different types of features; h(·) and g(·) are the output feature map of the double-bottleneck sub-convolution network and the feature map of the backbone network, respectively; i is the hierarchical index of the image pyramid and the backbone network; I0 and I i Represent the original image and the image of the i-th layer in the image pyramid respectively.

4. The multi-scale object detection method guided by image pyramid features according to claim 1, characterized in that: In step S4, the Focal loss function is as follows: Among them, p and p * are sample prediction value and sample true value respectively; α t ∈(0,1) is a weight factor introduced for class 1, 1-α t is to introduce a weight factor for class-1; (1-p) γ is the modulation coefficient; The position regression loss term is expressed as: Among them, L reg (t,t * ) is represented by a smooth L1 loss; t = {x, y, w, h} represents the bounding box position information of the sample prediction, where {x, y} represents the center coordinates of the bounding box, and {w, h} represents the width and height of the bounding box; t * is the sample label of t; The target detection loss function is expressed as: in, is the number of samples for classification, is the number of samples for regression; w is the number of batches of training images; k is the index of a single sample in each batch of training samples; λ is the loss balance term.

Images