Multi-layer semantic perception, distillation and semi-supervised cooperative training target detection method

Abstract

The invention discloses a multi-layer semantic perception, distillation and semi-supervised cooperative training target detection method, and the method specifically comprises the following steps: constructing a data set: extracting a first part of images from image data, marking the first part of images to construct a supervised target detection data set, and taking the remaining images as an unmarked image data set, the data volume of the unlabeled image data set is greater than that of the supervised target detection data set; teacher model optimization: performing supervision training on the teacher model on the supervised target detection data set; constructing a teacher model, and executing self-distillation training on the teacher model; pseudo labels are generated for the unlabeled images through a teacher model, and adaptive screening is carried out based on confidence distribution; mapping a pseudo label to a strong enhanced sample through enhanced geometric transformation, carrying out semi-supervised training by using the strong enhanced sample and the pseudo label, and introducing a feature layer distillation constraint at the same time; optimizing a student model; and outputting the target detection model obtained through training.

Description

Technical Field

[0001] This invention is applied to the field of artificial intelligence, specifically a multi-layer semantic perception, distillation and semi-supervised collaborative training object detection method. Background Technology

[0002] With the development of deep learning, object detection has made significant progress in various fields such as autonomous driving, video surveillance, and intelligent manufacturing. Benefiting from large amounts of labeled data and strong supervision signals, supervised learning methods such as Faster R-CNN and the YOLO series have achieved good detection performance on public datasets. However, in practical applications, obtaining a large number of high-quality labeled samples is costly, limiting the generalization ability and deployment efficiency of supervised learning models. Therefore, researchers have proposed semi-supervised object detection methods, aiming to improve model performance by jointly training with a small amount of labeled data and a large amount of unlabeled data, reducing the dependence on manual annotation.

[0003] Semi-supervised learning methods based on a teacher-student architecture have attracted widespread attention. These methods typically utilize a teacher model to generate pseudo-labels on unlabeled data, which are then used by the student model for supervised learning. To enhance the credibility of the pseudo-labels, some methods introduce consistency constraint mechanisms, such as Consistent Teacher, which uses reasoning on different perturbation versions of the image to select pseudo-labels with high predictive consistency for training. These methods mitigate the impact of pseudo-label noise on training to some extent and improve the model's generalization ability.

[0004] However, existing technologies still have several shortcomings. First, mainstream pseudo-label selection methods largely rely on the class confidence and consistency judgments of the teacher model's output layer, ignoring intermediate semantic features and contextual structural information. This leads to the generation of low-quality or misleading pseudo-labels in complex scenes with target occlusion, density, or significant scale variations, thus affecting the training performance of student models. Second, teacher models often employ passive update strategies such as exponential moving average (EMA), which have limited expressive power and prevent continuous optimization of pseudo-label quality. Furthermore, student models often rely solely on pseudo-labels during training, lacking deep knowledge transfer from the teacher model, which hinders the model from learning more robust feature representations.

[0005] To address the aforementioned issues, this invention proposes a target detection method that integrates multi-layer semantic perception, knowledge distillation, and semi-supervised collaborative training. By introducing multi-layer semantic information and a context-aware mechanism to optimize pseudo-label selection, the accuracy and stability of pseudo-labels are significantly improved. Simultaneously, the teacher model employs a self-distillation and semi-supervised joint optimization strategy to enhance its adaptability to complex scenarios. During training, the student model not only learns using dynamic pseudo-labels but also performs distillation based on intermediate features and structural knowledge from the teacher model, thereby achieving bidirectional collaborative optimization between the teacher and student models. This approach effectively improves the accuracy and robustness of target detection on low-label samples, overcoming key problems in existing methods such as unstable pseudo-labels, limited teacher performance, and hindered student training. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a multi-layer semantic perception, distillation and semi-supervised collaborative training target detection method to address the shortcomings of the prior art.

[0007] To address the aforementioned technical problems, the present invention provides a multi-layer semantic perception, distillation, and semi-supervised collaborative training method for target detection, specifically comprising the following steps:

[0008] Building the dataset:

[0009] The first part of the images is extracted from the image data and labeled to construct a supervised object detection dataset. The remaining images are used as an unlabeled image dataset, and the amount of data in the unlabeled image dataset is greater than the amount of data in the supervised object detection dataset.

[0010] Teacher model optimization:

[0011] The teacher model is trained under supervised supervision on the supervised object detection dataset.

[0012] Construct a teacher model and perform self-distillation training on the teacher model;

[0013] For unlabeled images, pseudo-labels are generated by the teacher model and adaptively filtered based on the confidence distribution. The pseudo-labels are mapped onto strongly enhanced samples through enhanced geometric transformation. Semi-supervised training is performed using the strongly enhanced samples and their pseudo-labels, while feature layer distillation constraints are introduced.

[0014] Student model optimization:

[0015] The teacher model parameters obtained from teacher model optimization are frozen, and a lightweight object detection model is selected as the student model. The student model is trained by knowledge distillation on a supervised object detection dataset, and semi-supervised training is performed on an unlabeled image dataset using the same pseudo-label generation method as the teacher model optimization process. The student model training includes both pseudo-label supervision loss and distillation loss.

[0016] Output the trained target detection model.

[0017] As one possible implementation, the construction of the teacher model and the self-distillation training step specifically involves: constructing the teacher model based on the exponential moving average (EMA) and performing self-distillation training on the teacher model. The self-distillation training includes at least: feature distillation of the intermediate feature layer and result distillation of the detection head output.

[0018] As a possible implementation, further, pseudo-labels are generated for unlabeled images by the teacher model and adaptively filtered based on the confidence distribution; the pseudo-labels are mapped onto strongly enhanced samples through enhanced geometric transformation, and semi-supervised training is performed using the strongly enhanced samples and their pseudo-labels, while introducing feature layer distillation constraints. Specifically, the steps are as follows: generating pseudo-labels for unlabeled images and performing semi-supervised training: weak enhancement is performed on the unlabeled image and input into the teacher model to obtain predicted boxes. The confidence values ​​of the latest N predicted boxes are accumulated by category, and the Gaussian mixture model parameters are iteratively solved using the expectation-maximization (EM) algorithm to obtain the classification threshold. Predicted boxes higher than the classification threshold are used as pseudo-labels for the weakly enhanced images; then, strong enhancement is performed on the unlabeled image, and the coordinates of the pseudo-labels are inversely mapped from the weak enhancement space to the original image space and forward mapped to the strong enhancement space according to the affine transformation matrix of the strong enhancement record to obtain pseudo-labels for the strongly enhanced images; the teacher model is trained using the strongly enhanced images and their pseudo-labels, and feature distillation constraints are applied to the intermediate feature layers during training.

[0019] As one possible implementation, the teacher model is further described as a two-stage object detection network or a single-stage object detection network, including any one of Faster R-CNN, YOLO, or DETR.

[0020] As one possible implementation, the intermediate feature layer distillation further includes normalizing the features at the channel level, using KL divergence to measure the difference between the teacher model and the student model on the selected set of intermediate feature layers, and forming the feature distillation loss.

[0021] As one possible implementation, the detection head output distillation further includes categorical distillation and regression distillation, wherein categorical distillation uses varifocal loss and regression distillation uses Smooth L1 loss.

[0022] As one possible implementation, further, weak enhancement includes resizing and / or flipping; strong enhancement includes random brightness and / or random deformation.

[0023] As one possible implementation, N is the cumulative number of most recently predicted box confidence scores by category, and N is 100.

[0024] As one possible implementation, a classification threshold is further used to distinguish between positive and negative samples, and predicted boxes above the classification threshold are used as pseudo-labels.

[0025] As one possible implementation, the pseudo-label coordinate mapping further includes at least: inversely mapping the pseudo-label coordinates of the weakly enhanced image according to the affine transformation matrix to restore the original image space, and then forward mapping the original image space coordinates to obtain the pseudo-label coordinates of the strongly enhanced image space.

[0026] The present invention adopts the above technical solution and has the following beneficial effects:

[0027] The pseudo-labels are more accurate and stable: After the teacher model generates candidate boxes for unlabeled samples under weak enhancement, the confidence scores are accumulated by category and the Gaussian mixture model is fitted by the EM algorithm to obtain an adaptive classification threshold to filter pseudo-labels. Compared with fixed threshold / simple consistency judgment, it can reduce the problem of mis-screening and noisy pseudo-labels in complex scenes (occlusion, density, scale changes).

[0028] Strong augmentation training provides more reliable label alignment: By utilizing the affine transformation matrix recorded during the augmentation process, the pseudo-label coordinates are inversely mapped from the weak augmentation space to the original image and then forward mapped to the strong augmentation space. This ensures the geometric consistency of the supervision signal for strongly augmented samples, reduces pseudo-label drift caused by strong geometric transformations, and thus improves the stability and robustness of semi-supervised training.

[0029] The teacher model's capabilities can be continuously enhanced, and the quality of pseudo-labels can be iteratively improved: The teacher model not only passively updates the EMA, but also overcomes the shortcomings of limited teacher performance improvement and difficulty in continuously optimizing pseudo-label quality through joint optimization of "self-distillation + semi-supervision" (using pseudo-label supervision and superimposing feature distillation constraints on unlabeled strong enhancement branches).

[0030] Student models learn more fully and generalize better: Student training introduces pseudo-label supervision and deep knowledge distillation from teachers (distillation of intermediate feature layers and detection head layers). Compared with training methods that rely solely on pseudo-labels, this helps students obtain more robust feature representations and reduces the performance ceiling caused by pseudo-label noise.

[0031] Improving detection performance with low labeling costs: With a small amount of labeled data and a large amount of unlabeled data, the detection accuracy and robustness under low-labeled sample conditions can be effectively improved through teacher-student collaborative optimization.

[0032] Highly practical and widely applicable: The algorithm is practical, low-cost, and can be reused in any target detection network, achieving simultaneous improvement in detection accuracy and robustness, making it easy to be quickly implemented in different detection frameworks and application scenarios. Attached Figure Description

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0034] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0036] like Figure 1 As shown, the present invention provides a multi-layer semantic perception, distillation, and semi-supervised collaborative training method for target detection, specifically including the following steps:

[0037] Building the dataset:

[0038] The first part of the images is extracted from the image data and labeled to construct a supervised object detection dataset. The remaining images are used as an unlabeled image dataset, and the amount of data in the unlabeled image dataset is greater than the amount of data in the supervised object detection dataset.

[0039] Teacher model optimization:

[0040] The teacher model is trained under supervised supervision on the supervised object detection dataset.

[0041] Construct a teacher model and perform self-distillation training on the teacher model;

[0042] For unlabeled images, pseudo-labels are generated by the teacher model and adaptively filtered based on the confidence distribution. The pseudo-labels are mapped onto strongly enhanced samples through enhanced geometric transformation. Semi-supervised training is performed using the strongly enhanced samples and their pseudo-labels, while feature layer distillation constraints are introduced.

[0043] Student model optimization:

[0044] The teacher model parameters obtained from teacher model optimization are frozen, and a lightweight object detection model is selected as the student model. The student model is trained by knowledge distillation on a supervised object detection dataset, and semi-supervised training is performed on an unlabeled image dataset using the same pseudo-label generation method as the teacher model optimization process. The student model training includes both pseudo-label supervision loss and distillation loss.

[0045] Output the trained target detection model.

[0046] As one possible implementation, the construction of the teacher model and the self-distillation training step specifically involves: constructing the teacher model based on the exponential moving average (EMA) and performing self-distillation training on the teacher model. The self-distillation training includes at least: feature distillation of the intermediate feature layer and result distillation of the detection head output.

[0047] As a possible implementation, further, pseudo-labels are generated for unlabeled images by the teacher model and adaptively filtered based on the confidence distribution; the pseudo-labels are mapped onto strongly enhanced samples through enhanced geometric transformation, and semi-supervised training is performed using the strongly enhanced samples and their pseudo-labels, while introducing feature layer distillation constraints. Specifically, the steps are as follows: generating pseudo-labels for unlabeled images and performing semi-supervised training: weak enhancement is performed on the unlabeled image and input into the teacher model to obtain predicted boxes. The confidence values ​​of the latest N predicted boxes are accumulated by category, and the Gaussian mixture model parameters are iteratively solved using the expectation-maximization (EM) algorithm to obtain the classification threshold. Predicted boxes higher than the classification threshold are used as pseudo-labels for the weakly enhanced images; then, strong enhancement is performed on the unlabeled image, and the coordinates of the pseudo-labels are inversely mapped from the weak enhancement space to the original image space and forward mapped to the strong enhancement space according to the affine transformation matrix of the strong enhancement record to obtain pseudo-labels for the strongly enhanced images; the teacher model is trained using the strongly enhanced images and their pseudo-labels, and feature distillation constraints are applied to the intermediate feature layers during training.

[0048] As one possible implementation, the teacher model is further described as a two-stage object detection network or a single-stage object detection network, including any one of Faster R-CNN, YOLO, or DETR.

[0049] As one possible implementation, the intermediate feature layer distillation further includes normalizing the features at the channel level, using KL divergence to measure the difference between the teacher model and the student model on the selected set of intermediate feature layers, and forming the feature distillation loss.

[0050] As one possible implementation, the detection head output distillation further includes categorical distillation and regression distillation, wherein categorical distillation uses varifocal loss and regression distillation uses Smooth L1 loss.

[0051] As one possible implementation, further, weak enhancement includes resizing and / or flipping; strong enhancement includes random brightness and / or random deformation.

[0052] As one possible implementation, N is the cumulative number of most recently predicted box confidence scores by category, and N is 100.

[0053] As one possible implementation, a classification threshold is further used to distinguish between positive and negative samples, and predicted boxes above the classification threshold are used as pseudo-labels.

[0054] As one possible implementation, the pseudo-label coordinate mapping further includes at least: inversely mapping the pseudo-label coordinates of the weakly enhanced image according to the affine transformation matrix to restore the original image space, and then forward mapping the original image space coordinates to obtain the pseudo-label coordinates of the strongly enhanced image space.

[0055] Example:

[0056] A target detection method combining multi-layer semantic perception, knowledge distillation, and semi-supervised collaborative training specifically includes the following steps:

[0057] Step 1: Building the dataset

[0058] Collect a large amount of data, extract a small portion for annotation, and construct a supervised object detection dataset. The remaining data consists of a large-scale unlabeled image dataset, which is larger than the supervised object detection dataset.

[0059] Step Two: Teacher Model Optimization

[0060] Step 2.1 Standard supervised training

[0061] Initialize teacher model T θ In supervised object detection datasets Training is then performed. The teacher model can employ mainstream two-stage or single-stage detection architectures, such as Faster R-CNN, YOLO, and DETR. The following uses YOLOv11 as an example; YOLOv11-x can be used as the teacher model, and its loss function is typically... in The classification loss ensures that the model can correctly identify which category an object in an image belongs to. Bounding box regression loss is the part used to optimize the difference between the predicted bounding box and the true bounding box. It uses probabilistic regression to predict the box, allowing the network to quickly focus on the location distribution in the vicinity of the label location.

[0062] Step 2.2 Self-distillation optimization

[0063] To further enhance the teacher model's ability to integrate and generalize features from different semantic layers, a self-distillation mechanism is introduced. An exponential moving average (EMA) mechanism is used to construct the teacher model T1, and multi-level self-distillation optimization is performed on it, including distillation of intermediate feature layers and detection result layers. This improves the overall network's generalization ability and the quality of multi-layer semantic representation, and provides a solid foundation for subsequent high-quality pseudo-label generation.

[0064] Its loss function is as follows:

[0065]

[0066] in

[0067] Intermediate characteristic distillation loss Indicates channel level

[0068] Feature normalization allows focus to be placed on the most salient regions within each channel. KL represents the KL divergence, L represents the set of selected intermediate feature layers, and F... s 1 represents the output of the feature layer of the student model. This is the output of the feature layer of the teacher model. For the detector head output, the classification and regression results are distilled separately, with the classification part using varifocal loss (divergence loss). The regression part uses Smooth L1 loss. Let the predicted class distribution be P. s The corresponding teacher output is P. t The predicted bounding box is B. s The corresponding teacher output is B. t fg is the matching value between the predicted bounding box and the annotation box, and its formula is as follows:

[0069]

[0070]

[0071] α and γ are hyperparameters used to balance the weights of positive and negative samples.

[0072] Step 2.3 Self-distillation and semi-supervised optimization

[0073] Step 2.3.1 Pseudo-tag generation

[0074] After performing weak enhancements (such as resizing or flipping) on ​​the unlabeled image, it is input into the teacher model to obtain predicted bounding boxes. The confidence values ​​(which can be set to 100) of the latest N boxes are accumulated according to their categories. Then, the optimal parameters of the Gaussian mixture model are solved iteratively using the expectation-maximization (EM) algorithm. The classification threshold (distinguishing between positive and negative samples) is obtained through the Gaussian mixture model. Boxes with values ​​higher than the threshold are used as pseudo-labels g for this unlabeled weakly enhanced image. t1 The unlabeled image is then subjected to strong enhancement (e.g., random brightness, random deformation). Based on the affine transformation matrix recorded during image enhancement, the label coordinates are inversely mapped from the weak enhancement space to the original image space, and then orthogonally mapped back to the strong enhancement space, resulting in the pseudo-label g of the strongly enhanced image. t2

[0075] Step 2.3.2 Self-distillation and semi-supervised distillation

[0076] For supervised image branches, self-distillation optimization is performed, consistent with step 2.2; for unlabeled images, strongly enhanced images are input into the network, and the intermediate feature layers use feature distillation consistent with step 2.2. The output uses pseudo-labels g. t2 As labels, a loss function consistent with standard supervision is used.

[0077] Its total loss function is shown below:

[0078]

[0079] The final model is obtained after training. T

[0080] Step 3: Student Model Optimization

[0081] Step 3.1 Knowledge Distillation

[0082] Use the model obtained in step 2 T As the teacher model, the teacher model weights are frozen and not updated; a lightweight small model, such as YOLOv11-S, is selected as the student model. Multi-level self-distillation optimization is then performed on this model, including distillation of the intermediate feature layer and the detection result layer. The loss function is shown below:

[0083]

[0084] Step 3.2 Knowledge Distillation and Semi-Supervised Learning

[0085] Step 3.2.1 Pseudo-tag generation

[0086] After performing weak enhancements (such as resizing or flipping) on ​​the unlabeled image, it is input into the teacher model to obtain predicted bounding boxes. The confidence values ​​(which can be set to 100) of the latest N boxes are accumulated according to their categories. Then, the optimal parameters of the Gaussian mixture model are solved iteratively using the expectation-maximization (EM) algorithm. The classification threshold (distinguishing between positive and negative samples) is obtained through the Gaussian mixture model. Boxes with values ​​higher than the threshold are used as pseudo-labels g for this unlabeled weakly enhanced image. t1 The unlabeled image is then subjected to strong enhancement (e.g., random brightness, random deformation). Based on the affine transformation matrix recorded during image enhancement, the label coordinates are inversely mapped from the weak enhancement space to the original image space, and then orthogonally mapped back to the strong enhancement space, resulting in the pseudo-label g of the strongly enhanced image. t2

[0087] Step 3.3.2 Knowledge Distillation and Semi-Supervised Learning

[0088] For supervised image branches, knowledge distillation optimization is performed, consistent with step 3.1; for unlabeled images, strongly enhanced images are input into the network, and the intermediate feature layers use feature distillation consistent with step 2.2. The output uses pseudo-labels g. t2 As labels, a loss function consistent with standard supervision is used.

[0089] Its total loss function is shown below:

[0090]

[0091] The final model is obtained after training. s .

[0092] The above description is an embodiment of the present invention. For those skilled in the art, any equivalent changes, modifications, substitutions and variations made in accordance with the teachings of the present invention without departing from the principles and spirit of the present invention should be covered by the present invention.

Claims

1. A multi-layer semantic perception, distillation, and semi-supervised collaborative training method for target detection, characterized in that, Specifically, the steps include the following: Building the dataset: The first part of the images is extracted from the image data and labeled to construct a supervised object detection dataset. The remaining images are used as an unlabeled image dataset, and the amount of data in the unlabeled image dataset is greater than the amount of data in the supervised object detection dataset. Teacher model optimization: The teacher model is trained under supervised supervision on the supervised object detection dataset. Construct a teacher model and perform self-distillation training on the teacher model; For unlabeled images, pseudo-labels are generated by the teacher model and adaptively filtered based on the confidence distribution. The pseudo-labels are mapped onto strongly enhanced samples through enhanced geometric transformation. Semi-supervised training is performed using the strongly enhanced samples and their pseudo-labels, while feature layer distillation constraints are introduced. Student model optimization: The teacher model parameters obtained from teacher model optimization are frozen, and a lightweight object detection model is selected as the student model. The student model is trained by knowledge distillation on a supervised object detection dataset, and semi-supervised training is performed on an unlabeled image dataset using the same pseudo-label generation method as the teacher model optimization process. The student model training includes both pseudo-label supervision loss and distillation loss. Output the trained target detection model.

2. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 1, characterized in that: The steps of constructing a teacher model and performing self-distillation training on the teacher model are as follows: constructing a teacher model based on the exponential moving average (EMA) and performing self-distillation training on the teacher model. The self-distillation training includes at least: performing feature distillation on the intermediate feature layer and performing result distillation on the output of the detection head.

3. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 1, characterized in that: The process involves generating pseudo-labels for unlabeled images using a teacher model and adaptively filtering them based on confidence distribution. The pseudo-labels are then mapped onto strongly enhanced samples via enhanced geometric transformation. Semi-supervised training is performed using the strongly enhanced samples and their pseudo-labels, while introducing feature layer distillation constraints. Specifically, the process involves generating pseudo-labels for unlabeled images and performing semi-supervised training: Weak enhancement is applied to the unlabeled image, which is then input into the teacher model to obtain predicted bounding boxes. The confidence values ​​of the latest N predicted bounding boxes are accumulated by category, and the Gaussian mixture model parameters are iteratively solved using the expectation-maximization (EM) algorithm to obtain a classification threshold. Predicted bounding boxes with values ​​higher than the classification threshold are used as pseudo-labels for the weakly enhanced image. Strong enhancement is then applied to the unlabeled image, and the pseudo-label coordinates are inversely mapped from the weak enhancement space to the original image space and then forward mapped to the strong enhancement space based on the affine transformation matrix recorded by the strong enhancement, to obtain the pseudo-labels for the strongly enhanced image. The teacher model is trained using the strongly enhanced image and its pseudo-labels, and feature distillation constraints are applied to the intermediate feature layers during training.

4. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 1, characterized in that: The teacher model is a two-stage object detection network or a single-stage object detection network, including any one of Faster R-CNN, YOLO, or DETR.

5. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 2, characterized in that: The intermediate feature layer distillation involves normalizing the features at the channel level, then using KL divergence to measure the difference between the teacher model and the student model on the selected intermediate feature layer set, and forming the feature distillation loss.

6. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 2, characterized in that: The output distillation of the detection head includes categorical distillation and regression distillation, wherein categorical distillation uses varifocal loss and regression distillation uses Smooth L1 loss.

7. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 3, characterized in that: The weak enhancements include resize and / or flip; the strong enhancements include random brightness and / or random deformation.

8. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 3, characterized in that: N is the cumulative number of most recently predicted bounding box confidence scores by category, and N is 100.

9. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 3, characterized in that: The classification threshold is used to distinguish between positive and negative samples, and predicted boxes with values ​​higher than the classification threshold are used as pseudo-labels.

10. The multi-layer semantic perception, distillation, and semi-supervised collaborative training target detection method according to claim 3, characterized in that: The pseudo-label coordinate mapping includes at least: inversely mapping the pseudo-label coordinates of the weakly enhanced image according to the affine transformation matrix to restore the original image space, and then forward mapping the original image space coordinates to obtain the pseudo-label coordinates of the strongly enhanced image space.

Images