Marine three-dimensional air gun array configuration method based on machine learning algorithm

Abstract

The invention provides a marine three-dimensional air gun array configuration method based on a machine learning algorithm, and belongs to the technical field of substance or object detection, and the method comprises the steps: collecting historical input parameters, and carrying out the data preprocessing; in data preprocessing, transversely splicing the two-dimensional sequence features and the single-valued scalar features to form a high-dimensional feature vector; and establishing an original data set, constructing an initial model based on a machine learning algorithm, and applying the trained configuration model to the to-be-configured marine three-dimensional air gun array so as to output current configuration parameters. Through cross-type feature splicing in a preprocessing stage, high-dimensional feature fusion is realized, rapid adaptation of three-dimensional gun array configuration under different sea conditions can be realized in combination with a machine learning technology, and high-accuracy ocean three-dimensional air gun array configuration parameters can be rapidly obtained after training, so that operation steps are simplified, the computing resource demand is reduced, and the construction efficiency is improved. And the execution efficiency is high, and the accuracy meets actual engineering requirements.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of detection of substances or objects, and particularly relates to a marine stereoscopic air gun array configuration method based on a machine learning algorithm. BACKGROUND

[0002] In the field of offshore seismic exploration, a stereoscopic air gun array plays a crucial role as a seismic source device. The working principle of the stereoscopic air gun array is to generate a shock wave by releasing high-pressure gas to provide sufficient energy for seismic data acquisition. In actual operation, the sinking combination mode and trigger time coding scheme of the stereoscopic air gun array have a significant impact on the exploration effect. Different sinking combination modes and trigger time coding schemes of the subarray will lead to differences in far-field wavelets and frequency spectrum, thereby affecting the quality and accuracy of the exploration data.

[0003] However, the traditional simulation method has obvious shortcomings in dealing with these complex combinations. Specifically, because the stereoscopic air gun array configuration needs to consider many factors, the conventional simulation method is often time-consuming and inefficient, making it difficult to meet the urgent needs of field construction. Therefore, in previous work, the best sinking depth and best delay firing time coding scheme combination for the exploration target are mainly selected by manual experiments. Although the manual experience screening method can achieve the purpose to some extent, it still has the problem of low efficiency and is easily affected by human factors, resulting in fluctuations in the accuracy and reliability of the results.

[0004] In summary, further improvement is needed to overcome the shortcomings of the prior art and provide an efficient and accurate parameter configuration method for the optimization design of the air gun array in offshore seismic exploration. SUMMARY

[0005] The application provides a marine stereoscopic air gun array configuration method based on cross-type feature splicing and a machine learning algorithm to quickly and accurately output the marine stereoscopic air gun array configuration, thereby solving the above problems.

[0006] To achieve the above purpose, the technical solution adopted by the application is as follows: The application provides a marine stereoscopic air gun array configuration method based on a machine learning algorithm, including the following steps: Collecting historical input parameters representing the characteristics of the marine stereoscopic air gun array source wavelet, the historical input parameters including two-dimensional sequence features and single-value scalar features; Data preprocessing is performed on the historical input parameters; in the data preprocessing, at least interpolation or truncation operation is performed on the two-dimensional sequence features to unify the data length, and the two-dimensional sequence features with unified data length are transversely spliced with the single-value scalar features to form a high-dimensional feature vector; preferably, the transverse splicing is to arrange the two-dimensional sequence features with unified data length in the order of sub-wave sequence and frequency spectrum sequence, and transversely splice them with the single-value scalar features in sequence to form a 1xN-dimensional high-dimensional feature vector, N being the sum of the total dimension of the two-dimensional sequence features and the dimension of the single-value scalar features; the single-value scalar features are arranged in the order of initial peak value, peak-to-peak value, bubble ratio, frequency band minimum value, and frequency band maximum value. The historical configuration parameters corresponding to the historical input parameters are screened; The historical input parameters after data preprocessing and the corresponding historical configuration parameters are combined to form an original data set, and the original data set is divided into a training set and a reserved validation set; An initial model is constructed based on a machine learning algorithm, the initial model is trained through the training set, and a trained model is obtained; The trained model is verified and fine-tuned through the reserved validation set, and a configuration model is obtained; The current input parameters corresponding to the to-be-configured marine three-dimensional air gun array are collected; the current input parameters are subjected to the data preprocessing, and the current input parameters after data preprocessing are input into the configuration model to output the current configuration parameters of the to-be-configured marine three-dimensional air gun array.

[0007] Further, the machine learning algorithm used to construct the initial model is one of XGBoost algorithm, LightGBM algorithm, CatBoost algorithm, random forest algorithm, fully connected deep neural network algorithm, or one-dimensional convolutional neural network algorithm.

[0008] Further, in the training process of the initial model, cross-validation method is used for training, including the following steps: The training set is equally divided into z training subsets, and z-1 training subsets are used to train the initial model each time, and 1 training subset is used to verify the training effect, and the cross-validation is completed by repeating z times; z is a positive integer and is greater than or equal to 2.

[0009] Further, in the training process of the initial model, hyperparameter optimization is also performed, and the hyperparameter optimization adopts a grid search method or a Bayesian optimization method.

[0010] Further, the two-dimensional sequence features are sub-waves and frequency spectra, and the single-value scalar features at least include initial peak value, peak-to-peak value, bubble ratio, frequency band minimum value, and frequency band maximum value.

[0011] Further, the historical configuration parameters and the current configuration parameters are used for the depth of each subarray and the delay time between each subarray in the marine three-dimensional air gun array.

[0012] Further, when the two-dimensional sequence features are interpolated or truncated, the length of the unified two-dimensional sequence features is a preset fixed length. For the two-dimensional sequence features with a length less than the preset fixed length after interpolation, a preset value is used to fill the two-dimensional sequence features to the preset fixed length.

[0013] Further, when the original data set is divided into a training set and a reserved validation set, the division ratio of the training set to the reserved validation set is 1:1.

[0014] Further, the validation of the training model includes the following steps: Calculate the root mean square error and the mean absolute error of the validation result. Based on the root mean square error and the mean absolute error, the validation label of the configuration model is output by threshold comparison method.

[0015] Further, after outputting the current configuration parameters of the marine three-dimensional air gun array to be configured, the current configuration parameters are fine-tuned, and the amplitude of the fine-tuning is not more than 10% of the values of the parameters in the current configuration parameters.

[0016] Compared with the prior art, the present application has the following beneficial effects: The present application realizes high-dimensional feature fusion by cross-type feature splicing in the preprocessing stage, thereby maintaining the correlation between parameters in different fields, and combining machine learning technology to realize rapid adaptation of three-dimensional gun array configuration under different sea conditions. After training, the marine three-dimensional air gun array configuration parameters with high accuracy can be quickly obtained. Unlike conventional simulation or iterative optimization methods, the present application simplifies the operation steps, reduces the demand for computing resources, and has high execution efficiency and accuracy that meets the actual engineering requirements. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 The method flowchart in the specific embodiments of the present application. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] It should also be noted that, unless otherwise specified, the methods used in this application are conventional methods; and the raw materials and equipment used are, unless otherwise specified, conventional commercially available products.

[0022] This application proposes a method for configuring a marine three-dimensional air gun array based on machine learning algorithms, such as... Figure 1 As shown, the specific steps include: 1. Historical input parameter collection and historical configuration parameter filtering; 1.1 Historical Input Parameter Acquisition: Historical input parameters characterizing the wavelet characteristics of the marine three-dimensional airgun array source were acquired. These historical input parameters include two-dimensional sequence features and single-valued scalar features. The two-dimensional sequence features are wavelets and spectra. Specifically, the wavelets are obtained by denoising the original vibration signals acquired by hydrophones, and the spectra are generated by the wavelets through Fast Fourier Transform (FFT). The single-valued scalar features include at least the initial peak value (amplitude of the first wavelet peak, unit: barm), peak-to-peak value (difference between the positive and negative maximum amplitudes of the wavelet, unit: barm), wave-bubble ratio (the ratio of the energy of the main wavelet pulse to the energy of subsequent bubble pulses, with the energy of the first 10ms of the main pulse and the energy of the bubble pulses from 10ms to 50ms), minimum and maximum frequency band values ​​(the boundary of the frequency range where the spectral energy is ≥ 10% of the maximum energy, unit: Hz).

[0023] It should be noted that the sources of historical input parameters must cover existing data types in the field, including measured data from marine seismic exploration, such as vibration data collected by hydrophones in a 3D exploration of an oil field; and also numerical simulation data, such as simulation data of the finite element method based on wave theory. The sample size of a single batch must be no less than 1,000 sets to ensure the generalization ability of subsequent model training.

[0024] 1.2 History configuration parameter screening: screening the history configuration parameters corresponding to the above history input parameters, the history configuration parameters being the sinking depth of each subarray in the marine three-dimensional airgun array and the delay shooting time between each subarray. Among them, the sinking depth is usually selected as 5-15 meters, and 6-10 meters is preferred for shallow sea exploration; the delay shooting time is selected as 0-5 milliseconds to avoid wavelet interference, and the screening standard is the parameter combination that can meet the requirements of seismic data quality such as effective frequency bandwidth and high signal-to-noise ratio after being verified by practice or optimized by simulation.

[0025] 2. History input parameter preprocessing and original data set construction; 2.1 History input parameter preprocessing: data preprocessing is performed on the history input parameters, including the following steps: First, interpolation or truncation operation is performed on the two-dimensional sequence characteristics (wavelet, spectrum) to unify the length of the two-dimensional sequence characteristics to a preset fixed length; the preset fixed length can be selected according to the application scenario, and is usually set to 256, which corresponds to a wavelet time sequence covering 0-256 ms and a spectrum frequency range covering 0-200 Hz; the interpolation method can be selected according to the feature type, such as linear interpolation for wavelet sequence and cubic spline interpolation for spectrum; if the length of the two-dimensional sequence characteristics after interpolation is still less than the preset fixed length, the two-dimensional sequence characteristics are filled to the preset fixed length by a preset value, one example of which is NaN, which can be replaced by the mean value of the same batch of data sequences to avoid the influence of missing values on model training.

[0026] After that, the two-dimensional sequence characteristics with unified length are horizontally spliced with single-value scalar characteristics to form a high-dimensional feature vector; the horizontal splicing is to arrange the two-dimensional sequence characteristics with unified data length in the order of wavelet sequence and spectrum sequence, and horizontally splice them with single-value scalar characteristics in turn to form a high-dimensional feature vector of 1xN, N being the sum of the total dimension of the two-dimensional sequence characteristics and the dimension of the single-value scalar characteristics. Among them, the single-value scalar characteristics are arranged in the order of initial peak value, peak-to-peak value, bubble ratio, minimum frequency band value and maximum frequency band value.

[0027] Specifically, taking a single sample with a preset fixed length of 256 as an example, in the two-dimensional sequence characteristics, the wavelet is unified to 256 dimensions (time step 1 ms, covering 0-256 ms) after interpolation, such as [w1, w2,..., w 256 ]; the spectrum is unified to a 256-dimensional sequence (frequency step 0.78125 Hz, covering 0-200 Hz) after interpolation, such as [s1, s2,..., s 256];single value scalar is 6 dimensional (primary peak value 82.4 barm, peak to peak value 127.5 barm, bubble ratio 22.5, frequency band minimum value 6.42 Hz, frequency band maximum value 88.12 Hz, subwave energy 1500 J), as [p1, p2,..., p6]. After transverse splicing: [w1, w2,..., w 256 , s1, s2,..., s 256 , p1, p2,..., p6], that is, a 1x518 dimensional feature vector is formed, where the first 256 dimensions correspond to the amplitude of each time point of the subwave, the 257th-512th dimensions correspond to the energy of each frequency point of the spectrum, and the 513th-518th dimensions correspond to the 6 single value scalar parameters; the 1x518 dimensional feature vector contains energy features and other supplementary scalars outside the subwave / spectrum, each dimension corresponds to a feature value, there is no row / column dimension expansion, only the feature dimension is superimposed. At the multi-sample level, multiple single-sample high-dimensional feature vectors are stacked longitudinally to form a feature matrix MxN dimensional, M is the number of samples, and N is the single-sample feature dimension. Thus, the feature vectors of the above example 1000 samples are arranged longitudinally to form a 1000x518 dimensional feature matrix, which can be used as an input data set for model training.

[0028] 2.2 Original data set construction and division: the historical input parameters after data preprocessing and the corresponding historical configuration parameters are combined to form an original data set, and the storage of the original data set adopts the efficient HDF5 format in the field, which is convenient for reading and writing of large-scale sequence data; preferably, the original data set is randomly divided into a training set and a reserved validation set in a 1:1 ratio. Among them, the reserved validation set is only used for final model generalization ability evaluation and does not participate in the preliminary training and parameter optimization.

[0029] The training set is , represents the ith feature vector with m dimensions, which in the present application is an input parameter containing subwave, spectrum and primary peak value, peak to peak value, bubble ratio, frequency band minimum value, frequency band maximum value; represents the ith sample label, which in the present application is the corresponding sink depth and delay excitation time coding scheme.

[0030] 3. Initial model construction and training; 3.1 Initial model construction: an initial model is constructed based on a machine learning algorithm, and the machine learning algorithm used is one of an XGBoost algorithm, a LightGBM algorithm, a CatBoost algorithm, a random forest algorithm, a fully connected deep neural network algorithm, or a one-dimensional convolutional neural network algorithm. Among them, the XGBoost algorithm is preferred as the machine learning algorithm of the present application due to its advantages in processing structured data and complex nonlinear relationships. The specific implementation follows the gradient boosting decision tree (GBDT) improvement framework, specifically: the XGBoost model realizes the mapping of input features-configuration parameters by integrating multiple CART regression trees, each tree is trained based on the prediction residual of the previous tree, and the final output is the sum of the prediction values of all trees. Among them, the residual = historical configuration parameter true value - prediction value of the previous N trees.

[0031] Specifically, the initial model contains K trees, and K of the present application is the number of CART regression trees allocated by the model according to the amount of training set data. Then the single tree of the model is defined as: ; The above formula can be simplified as: ; In the formula, is the analysis value of the i-th sample; is the output of the i-th CART regression tree; represents the K-th decision tree, and the sum of the analysis values of each tree is calculated as the final analysis value; is the sum of the analysis values of all K decision trees at present, and the objective function is set as: ; In the formula, the objective function obj is composed of two parts. The first term is the loss function used to evaluate the loss or error between the model prediction value and the true value, and the second term is the regularization term used to control overfitting.

[0032] 3.2 Initial model training: the initial model is trained through the training set, and the specific process is as follows: First, cross-validation training is used: the training set is evenly divided into z training subsets, z is a positive integer and greater than or equal to 2; the present application takes z = 5, that is, five-fold cross-validation, and each time 4 training subsets are used to train the initial model and 1 training subset is used to verify the training effect, and the cross-validation is repeated 5 times. This method can make full use of limited data and reduce the risk of overfitting.

[0033] The hyperparameter optimization is performed simultaneously during the training process, and the grid search method or the Bayesian optimization method is used for the hyperparameter optimization. In the hyperparameter grid and setting, the key hyperparameters to be optimized by the XGBoost model are defined, and a possible value range is set for the key hyperparameters in the form of a dictionary, including: learning_rate (learning rate / step size); max_depth (maximum depth of the tree); min_child_weight (minimum leaf node weight); subsample / colsample_bytree (sample / feature sampling ratio); reg_alpha, reg_lambda (L1 and L2 regularization parameters); Grid search and hyperparameter optimization: perform grid search, for each set of hyperparameters, perform the above five-fold cross-validation training, and use the average performance indicator (such as mean square error) of the five-fold validation as the evaluation basis for the pros and cons of the parameter combination. Finally, the average performance optimal hyperparameter combination is selected as the optimal configuration of the model, and the final XGBoost model is retrained on the entire training set (excluding the independent validation set).

[0034] The training process is based on the gradient boosting algorithm, which iteratively constructs multiple decision trees, each tree learns and corrects the prediction residual of the previous tree, and finally the weighted sum of the prediction results of all trees is taken as the final output.

[0035] In addition, during the training process, each sample in the training set The input parameters of the subwave, frequency spectrum, initial peak value, peak-to-peak value, bubble ratio, frequency band minimum value, and frequency band maximum value correspond to the encoding scheme of the sink depth and delay excitation time, which is used to calculate the first-order gradient and second-order gradient of the loss function; the formulas are as follows: ; ; In the formula, is the prediction value of the s-1th iteration, which realizes the efficient optimization of the objective function.

[0036] The model training termination condition is satisfied when any of the following conditions is met: (1) The number of iterations reaches the preset value of the hyperparameter n_estimators (number of iterations); (2) The MSE of the cross-validation for five consecutive rounds does not decrease; (3) The prediction residual of a single tree is less than 0.01.

[0037] After the training is completed, the trained model is obtained, and for subsequent deployment, the trained model can be temporarily stored, and after verification, it is serialized and saved as a configuration model in Joblib format.

[0038] 4. Model verification and configuration model acquisition; The trained model is verified, and the specific steps are as follows: First, the input parameters of the reserved verification set are processed according to the preprocessing method of step 2.1, and then input into the configuration model to obtain the predicted configuration parameters corresponding to the verification set; Second, the root mean square error (RMSE) and mean absolute error (MAE) of the verification result are calculated. Based on the two error indicators, the verification label of the configuration model is output by threshold comparison method. Further, the threshold is set according to the accuracy requirement of marine stereo air gun array configuration, which is usually: the RMSE of the sink depth is <0.5, the MAE is <0.3, the RMSE of the delay firing time is <0.1, and the MAE is <0.05.

[0039] The verification label includes qualified label or unqualified label. If the verification result does not meet the threshold, the hyperparameters need to be adjusted (such as expanding the grid search range), supplementing high-quality samples (such as adding new measured data across oilfields) or optimizing the preprocessing process (such as replacing the interpolation method) to meet the requirements; if the verification is qualified, the trained model is fine-tuned according to the verification result, such as adjusting the leaf node weight of a few trees, and finally the configuration model is obtained.

[0040] 5. Current configuration parameter output and fine-tuning; 5.1 Current input parameter acquisition and preprocessing: collect the current input parameters corresponding to the marine stereo air gun array to be configured, and the collection equipment and data type should be consistent with the historical input parameters in step 1.1; the current input parameters are preprocessed in the same way as the historical input parameters in step 2.1, ensuring that the input feature format matches the input requirements of the configuration model.

[0041] 5.2 Current configuration parameter output: input the preprocessed current input parameters into the configuration model to output the current configuration parameters of the marine stereo air gun array to be configured, i.e. the sink depth of each sub-array and the delay firing time between each sub-array.

[0042] 5.3 Current configuration parameter fine-tuning: after outputting the current configuration parameters of the marine stereoscopic airgun array to be configured, the current configuration parameters need to be fine-tuned in combination with the field construction conditions, such as the seabed topography, equipment load limit, etc. The fine-tuning range is not more than 10% of the value of each parameter in the current configuration parameters; for example: if the predicted sinking depth is 9 meters, but the field measurement shows that the seabed silt layer is relatively thick, it can be fine-tuned to 8.5 meters (fine-tuning range 5.6% <10%), to ensure construction safety and effect. Embodiment

[0043] Taking the preferred scheme of the airgun array as the source in a certain voyage as an example, the actual input parameters and the actual preferred configuration parameters are shown in Table 1: Table 1 Actual input parameters and actual preferred configuration parameters The input parameters are brought into the configuration model of the present embodiment, and the configuration parameters output by the model, i.e. the sinking depth and the delay firing time, are obtained, and the comprehensive accuracy obtained by comparing the configuration parameters output by the model with the actual preferred configuration parameters is shown in Table 2: Table 2 Sinking depth and delay firing analysis and comprehensive accuracy The average comprehensive accuracy in the present embodiment reaches 93% in the existing examples. From the comprehensive accuracy, it can be seen that the method of the present embodiment can reach the actual application level.

[0044] Finally, it should be noted that the above content is only used to illustrate the technical solutions of the present application, and is not a limitation on the protection scope of the present application. Simple modifications or equivalent replacements of the technical solutions of the present application made by those skilled in the art do not deviate from the essence and scope of the technical solutions of the present application.

Claims

1. A method for configuring a marine three-dimensional air gun array based on machine learning algorithms, characterized in that, Includes the following steps: Historical input parameters characterizing the wavelet characteristics of the marine three-dimensional air gun array source were collected. The historical input parameters included two-dimensional sequence features and single-valued scalar features. The historical input parameters are preprocessed; in the data preprocessing, at least the two-dimensional sequence features are interpolated or truncated to unify the data length, and the two-dimensional sequence features with the unified data length are horizontally concatenated with the single-value scalar features to form a high-dimensional feature vector. Filter the historical configuration parameters corresponding to the historical input parameters; The preprocessed historical input parameters and the corresponding historical configuration parameters are combined to form the original dataset, and the original dataset is divided into a training set and a reserved validation set. An initial model is built based on machine learning algorithms; The initial model is trained using the training set to obtain a trained model; The trained model is validated and fine-tuned using the reserved validation set to obtain the configured model; Collect the current input parameters corresponding to the marine three-dimensional air gun array to be configured; The current input parameters are preprocessed, and the preprocessed current input parameters are input into the configuration model to output the current configuration parameters of the marine three-dimensional air gun array to be configured.

2. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, The machine learning algorithm used to construct the initial model is one of the following: XGBoost algorithm, LightGBM algorithm, CatBoost algorithm, random forest algorithm, fully connected deep neural network algorithm, or one-dimensional convolutional neural network algorithm.

3. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, During the training of the initial model, cross-validation is used, including the following steps: The training set is divided into z training subsets. Each time, z-1 training subsets are used to train the initial model, and one training subset is used to verify the training effect. This process is repeated z times to complete the cross-validation. z is a positive integer and is greater than or equal to 2.

4. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, During the training of the initial model, hyperparameter optimization is also performed, which employs either a grid search method or a Bayesian optimization method.

5. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, The two-dimensional sequence features are wavelets and spectrum, and the single-valued scalar features include at least the initial peak value, peak-to-peak value, wave-to-bulb ratio, minimum frequency band value, and maximum frequency band value.

6. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, The historical configuration parameters and the current configuration parameters are the sinking depth of each subarray in the marine three-dimensional air gun array and the delay excitation time between each subarray.

7. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, When interpolating or truncating the two-dimensional sequence features, the length of the unified two-dimensional sequence features is a preset fixed length; For two-dimensional sequence features whose length after interpolation is less than the preset fixed length, the preset value is used to fill the preset fixed length.

8. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, When the original dataset is divided into a training set and a reserved validation set, the ratio of the training set to the reserved validation set is 1:

1.

9. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, Validating the trained model includes the following steps: Calculate the root mean square error and mean absolute error of the verification results; Based on the root mean square error and the mean absolute error, the verification label of the configuration model is output using a threshold comparison method.

10. The method for configuring a marine three-dimensional air gun array based on a machine learning algorithm according to claim 1, characterized in that, After outputting the current configuration parameters of the marine three-dimensional air gun array to be configured, the current configuration parameters are further fine-tuned, and the fine-tuning range does not exceed 10% of the values ​​of each parameter in the current configuration parameters.

Images