ADC global calibration method and circuit based on reinforcement learning

Abstract

The invention belongs to the field of ADC digital calibration algorithms, and particularly relates to an ADC global calibration method and circuit based on reinforcement learning, and the method comprises the steps: constructing a calibration network, training the calibration network through employing a reinforcement learning algorithm, inputting a to-be-calibrated ADC output signal into the trained calibration network, outputting a calibration amount through the calibration network, and carrying out the calibration of the to-be-calibrated ADC. And adding the calibration quantity to the ADC output signal to be calibrated to obtain a calibrated signal. On one hand, the method is suitable for various ADC structures, does not need to pay attention to error sources and nonlinear influences, and has the advantages of being high in universality and easy to transplant; and on the other hand, based on reinforcement learning, the final index is directly used as the target quantity to optimize the calibration network, so that the final calibration effect and the ideal value are optimally matched, and the final calibration capability is improved.

Description

Technical Field

[0001] The present invention belongs to the field of ADC digital calibration algorithms, and in particular relates to an ADC global calibration method and circuit based on reinforcement learning. Background Art

[0002] The conversion process of the analog-to-digital converter ADC includes four key stages: the first is the sampling stage, which is responsible for discretizing the continuous analog signal to form a periodic analog signal sample; the second is the holding stage, which ensures that the sampled signal value remains stable before the next sampling cycle for subsequent processing. These two stages are usually implemented by a dedicated sampling and holding circuit. This is followed by the quantization and encoding stages, which are responsible for mapping the discretized analog signal to a digital signal.

[0003] During the conversion process, the ADC may introduce errors due to factors such as device performance, manufacturing tolerances, and environmental variables. The type and size of these errors vary depending on the type and design of the ADC. Traditional calibration techniques are often only targeted at specific types of ADCs, with a limited range of errors that can be corrected, and the algorithm implementation is relatively complex, the hardware cost is high, and it is difficult to adapt to different platforms. In addition, these calibration methods may add additional circuit design to the analog part, which not only increases the complexity of the analog circuit design, but may also introduce additional performance uncertainties.

[0004] Neural network-based calibration methods have been proposed in recent years and have achieved good results, but most of them are based on supervised learning and rely on label true values. Due to the fitting error of the neural network, the final ADC index after supervised learning training and calibration may be far from the ideal ADC index, lacking global awareness. Summary of the invention

[0005] In view of the problems that traditional calibration algorithms are suitable for ADC structures and have single errors, and that neural network calibration algorithms rely on supervised training while ignoring final indicators, the present invention proposes an ADC global calibration method based on reinforcement learning, including: constructing a calibration network, and using a reinforcement learning algorithm to train the calibration network, inputting the output signal of the ADC to be calibrated into the trained calibration network, the calibration network outputting a calibration amount, and adding the calibration amount to the output signal of the ADC to be calibrated to obtain a calibrated signal.

[0006] Furthermore, the calibration network processes the ADC output signal to be calibrated including:

[0007] The ADC output signal to be calibrated is converted into a low-dimensional floating-point vector using a one-dimensional convolutional encoder;

[0008] The low-dimensional floating-point vector is subjected to feature extraction and feature fusion through multiple cascaded three-branch modules;

[0009] The signal output by the last three-branch module is mapped through a linear layer to obtain the calibration value;

[0010] Among them, each three-branch module processes the output signal through the residual branch, linear branch and nonlinear branch respectively, and then fuses the processed data through weighted summation, and uses the fused value as the output of the three-branch module.

[0011] Furthermore, the one-dimensional convolutional encoder is composed of cascaded one-dimensional convolutional layers and linear layers.

[0012] Furthermore, the signal processing process of the three-branch module is expressed as:

[0013] Y=αr×X+αl×Linear(X)+αn×ReLU(Linear(ReLU(Linear(X))))

[0014] Among them, Y represents the output of the three-branch module, X represents the input of the three-branch module; α r represents the fusion weight of the residual branch, α l represents the fusion weight of the linear branch, α n Represents the fusion weight of the nonlinear branch; Linear(·) represents the linear function; ReLU(·) represents the ReLU activation function layer.

[0015] Furthermore, in the process of training the calibration network using the reinforcement learning algorithm, the ADC output signal to be calibrated is used as the state value, and the calibration amount corresponding to the ADC output signal to be calibrated is used as the action value. The calibration network is used to select the action value with the largest reward value according to the state value, and the reward value is the spurious-free dynamic range (SFDR) of the ADC output signal to be calibrated.

[0016] Furthermore, the calculation of the multi-tone spurious free dynamic range of the ADC output signal to be calibrated includes: setting a range threshold β, taking the mean of other frequency points of the spectrum diagram of the ADC output signal to be calibrated except the basic frequency, and negating the sum of all frequency points in other frequency points that are greater than the mean + β to obtain the multi-tone spurious free dynamic range of the ADC output signal to be calibrated. The present invention also proposes an ADC global calibration circuit based on reinforcement learning, the circuit includes a control network and a calibration network, the control network includes an input FIFO buffer module, a neural controller module, a weight RAM module and a data output module, the calibration network calibrates the data input by the input FIFO buffer module and adds the data input by the input FIFO buffer module as the calibrated value to input the data output module, wherein:

[0017] An input FIFO buffer module is used to buffer the ADC output signal to be calibrated, and control the data buffered in the input FIFO buffer module to be input into the calibration network through the buffer output control signal of the neural controller module;

[0018] A neural controller module is used to control the read enable, write enable, read address and write address functions of the weight RAM module, and is also used for signal control;

[0019] Weight RAM module, used to store weight data needed to calibrate the network;

[0020] The data output module is used to ensure that the calibrated signal output by the calibration network is output only when the calibration enable and the calibration network output are both activated externally.

[0021] Furthermore, the linear branch includes a linear layer, and the nonlinear branch includes a linear layer and a ReLU activation function layer, wherein the linear layer uses a systolic array to implement N×K matrix transformation, each processing element in the systolic array multiplies the weight by the input, and passes the data together with the partial summation to the next processing element, and the matrix multiplication output at the end of each row is integrated through a multiplexer and output in a pipeline manner; the ReLU activation function layer uses combinational logic to implement symbol determination.

[0022] Furthermore, the one-dimensional convolutional encoder consists of a cascaded one-dimensional convolutional layer and a linear layer. The linear layer is implemented using weight rotation and a systolic array structure. Each processing element in the systolic array obtains the weight value from the weight RAM module. The intermediate layer results are stored in neuron registers and systematically rotated for matrix multiplication to maintain the systolic array structure. After one rotation cycle, the final result is output.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The present invention can be widely applied to various ADC architectures without the need for a calibration scheme for specific error type transformations. On the other hand, the present invention is based on reinforcement learning and directly uses the final indicator as the target quantity to optimize the calibration network, so that the final calibration effect and the ideal value are optimally matched, thereby improving the final calibration capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a schematic diagram of the overall architecture of the ADC calibration method based on reinforcement learning proposed in the present invention;

[0026] Figure 2 Schematic diagram of the reward function SFDRm in the MDP proposed by the present invention;

[0027] Figure 3 A schematic diagram of a customized neural network architecture for calibration proposed in the present invention;

[0028] Figure 4 A circuit design block diagram of the calibration network proposed in the present invention;

[0029] Figure 5 Spectrum diagrams before and after calibration in a specific example of the present invention;

[0030] Figure 6 This is a diagram showing the result of calibrating a multi-amplitude signal in a specific example of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] The present invention proposes an ADC global calibration method based on reinforcement learning, comprising: constructing a calibration network, and adopting a reinforcement learning algorithm to train the calibration network, inputting an ADC output signal to be calibrated into the trained calibration network, the calibration network outputting a calibration amount, and adding the calibration amount to the ADC output signal to be calibrated to obtain a calibrated signal.

[0033] As an optional implementation, the present invention provides an ADC global calibration method based on reinforcement learning, which mainly includes the following steps:

[0034] S1. Construct an ADC simulation model with multiple errors and structures. The model inputs time series and signals and outputs data with errors. At the same time, collect multiple sets of ADC measured data, covering multiple frequency bands.

[0035] S2. Normalize the ADC calibration task into a Markov decision process MDP: A Markov decision process MDP is characterized by a set of five basic elements, including: a state set S, which contains a finite number of states, representing various possible situations of the system; an action set A, which contains a finite number of actions, representing the behaviors that the agent can take in each state; a state transition probability matrix P, which describes the probability of the system transferring to another state given the current state and the selected action; a reward function R, which is used to evaluate the immediate reward obtained after taking a specific action in a specific state; and a discount factor γ, which is used to adjust the importance of future rewards relative to immediate rewards. Among them, the state set S, the action set A, the state transition probability matrix P, the reward function R and the discount factor γ together constitute the MDP quintuple (S, A, P, R, γ), which is used to formally describe and solve decision-making problems in uncertain environments. In this embodiment, the Markov decision process is applied to the ADC calibration problem, where the state s is the ADC output value with error (the amount to be calibrated), the action a is the calibration amount, and the reward function R is the final indicator of the ADC. This embodiment uses the spurious free dynamic range SFDR of the signal as the reward function. Specifically, in the present invention, the state value is s, and its state value selects an action a. When the action is executed, a new state value is obtained. According to the new state value, that is, the ADC output signal obtained after compensation for the calibration amount, the spurious free dynamic range SFDR of the signal is calculated as the reward value for this time.

[0036] S3. Design of calibration network structure: The present invention customizes and designs a new neural network structure named residual multilayer perceptron (ResMLP) as a calibration network, which is specifically used for calibration tasks. The calibration network processes the ADC output signal to be calibrated including:

[0037] First, a one-dimensional convolutional encoder is used to convert the raw binary output of the ADC into a low-dimensional floating-point vector to capture subtle errors in the data.

[0038] Secondly, the three-branch module in the calibration network further distributes the signal to three branches: a residual branch, a linear branch, and a nonlinear branch. Each branch processes the signal through a different activation function, and integrates the outputs of each branch through a weighted sum operation. Specifically, the input signal is not processed in any way in the residual branch, and only one linear layer is used in the linear branch to linearly process the input signal. In the nonlinear layer, a linear layer and a nonlinear activation function layer (in this embodiment, the nonlinear activation function selects the ReLU function) are used alternately to process the input signal. As a preferred implementation method, in the nonlinear branch, a linear layer and a nonlinear activation function layer are used alternately, and the output of the three-branch module can be expressed as:

[0039] Y=αr×X+αl×Linear(X)+αn×ReLU(Linear(ReLU(Linear(X))))

[0040] Among them, Y represents the output of the three-branch module, X represents the input of the three-branch module; α r represents the fusion weight of the residual branch, α l represents the fusion weight of the linear branch, α n Represents the fusion weight of the nonlinear branch; Linear(·) represents the linear function; ReLU(·) represents the ReLU activation function layer;

[0041] Finally, in order to further improve the performance and robustness of the model, the architecture repeats this multi-branch block after the weighted summation of the outputs of the three branches, that is, multiple three-branch module cascades (in this embodiment, two three-branch module cascades) are implemented. This iterative structure allows more detailed adjustment and fine-tuning and ultimately outputs a correction value as the final calibrated output.

[0042] S4. Solve the above MDP using reinforcement learning algorithm to obtain the optimal strategy, i.e., the optimal calibration network: Given the MDP, the reinforcement learning algorithm can find the optimal strategy (a network that maps state s to action a), which can enable the agent to obtain the maximum reward. In the ADC calibration problem, the strategy is the calibration network, which maps state s (the amount to be calibrated) to action a (calibration amount). Through continuous training of reinforcement learning, the calibration network can obtain the maximum reward, i.e., achieve the optimal SFDR, so that the ADC error can be calibrated and has the ability to calibrate multi-error and multi-structure ADCs.

[0043] The overall architecture of the ADC calibration method based on reinforcement learning in this embodiment is as follows: Figure 1 As shown. The calibration network interacts with the MDP environment, obtains the state s, the reward R, and outputs the calibration amount. The reinforcement learning algorithm uses the reward R to train the calibration network to obtain the maximum reward and. During the test, the error signal is collected online and then input into the calibration network for calibration. The specific implementation plan is as follows:

[0044] B1. Construct various structural ADC simulation models with various errors, including capacitor mismatch error, operational amplifier limited bandwidth error, memory effect, aperture jitter error, inter-channel gain mismatch error, and time mismatch error. ADC structures include Pipelined, SAR, Sigma-Delta, and Time-Interval. The ADC sampling frequency is 6G and the resolution is 12bit. For the input signal, this embodiment uses single / dual / multi-tone signals with a frequency in the range of [40MHz, 3GMHZ].

[0045] B2, such as Figure 1As shown in the upper part, this embodiment uses a reinforcement learning algorithm PPO to solve the MDP. For the ADC calibration task, the MDP is specifically set as follows: state s is defined as the original output of the ADC with error. Action a is defined as the calibration amount. The calibrated ADC output is equivalent to the original ADC output plus this calibration amount. The definition of the reward function R is the most critical. In view of the fact that the goal of this embodiment is to maximize the performance of the ADC, this embodiment defines the reward as a function closely related to the performance index. Since the signal-to-noise ratio and distortion ratio (SNDR) takes noise into account, it may lead to noise filtering rather than focusing on error calibration, so this embodiment does not use it as an optimization indicator. In addition, for high-speed applications, the spurious free dynamic range is more important, so this embodiment uses SFDR as a reward function. In addition, the method of the present invention is more flexible and allows the reward function to be adjusted as needed to focus on other indicators.

[0046] The original definition of SFDR is mainly applicable to single-tone signals, which limits its practicality in more complex scenarios. To address this limitation, this embodiment extends the SFDR indicator to multi-tone signals to make it more relevant and effective. The specific calculation method of multi-tone SFDR (expressed as SFDRm) is as follows: Figure 2 As shown, in this embodiment, the frequency point values ​​in the spectrum graph that are greater than a certain range of the mean value (in Figure 2 By setting a range threshold β, the negative sum of the frequencies of the frequency points in the spectrum graph other than the fundamental frequencies (fundamentalfrequencies) that are greater than the mean + β is taken as the multi-tone SFDR. Figure 2 The spectrum diagram can be obtained as follows: SFDRm = -(f1 + f2 + f3 + f4 + f5) This reward function setting method encourages the calibration network to reduce the spurious in the spectrum.

[0047] B3, such as Figure 3 The neural network architecture is designed as shown. The network design first adopts a one-dimensional convolutional encoder (BinaryConvolutional Encoder) as the main channel for preprocessing input data. The encoder consists of a cascaded one-dimensional convolutional layer (Comv1D) and a linear layer (Linear). The encoder is crucial for feature extraction, and it prepares the data for further processing. Traditional floating-point representations are difficult to capture small errors in input values. Therefore, this embodiment uses the binary code output by the original ADC as the input of the convolutional encoder. These binary codes are regarded as multi-channel one-dimensional time series, where the number of channels corresponds to the number of bits of the ADC. By applying the convolutional encoder, each data point is converted from a binary code to a low-dimensional floating-point vector. This binary convolutional encoder is designed to be sensitive to small errors in the ADC output, enabling it to capture useful information from the data.

[0048] The output signal from the 1D convolutional encoder is then processed by a three-branch module (ResMLP-Block), where the input signal is strategically distributed to three specialized branches: residual branch, linear branch, and non-linear branch. Each branch has a unique purpose and contributes to the overall performance of the network. The linear branch is equipped with one fully connected layer (or called linear layer), while the non-linear branch contains two fully connected layers. The main difference between the two branches lies in the nature of their activation functions. The linear branch adopts a linear activation function to preserve the original characteristics of the data, while the non-linear branch adopts a non-linear activation function to capture the complex relationships within the data.

[0049] These three branches do not operate in isolation, they are integrated into the network through a weighted sum operation. This integration is controlled by three learnable parameters that are adaptively updated during the training phase. Fine-tuning of these parameters provides an additional layer of optimization, thereby enhancing the network's ability to calibrate noisy data.

[0050] To further improve the performance and robustness of the model, the architecture repeats this multi-branch block after weighted summing of the outputs of the three branches. This iterative structure allows for more detailed adjustments and fine-tuning. The network ultimately outputs a correction value, which is added to the original ADC output to form the calibrated value.

[0051] B4. Use reinforcement learning to train the neural network, with 300 training rounds, MSE loss function, and learning rate of 0.01. Before starting training, the network parameters are randomly initialized with normal distribution.

[0052] B5. Figure 4 The block diagram of the designed calibration circuit is shown in Figure 1. The neural controller module Neural_Controller is responsible for weight reading and writing and signal control. It includes the following signals: read enable RD_EN, write enable WT_EN, read address RD_ADDR and write address WT_ADDR, which are used to enable reading, enable writing and addressing of read and write weights respectively. The buffer output control signal STR_EN controls the buffer output.

[0053] The input FIFO buffer module INPUT_FIFO_BUFFER stores incoming data while waiting for the weight write to complete, and if new weights need to be written to RAM. Controlled by the controller's buffer output control signal STR_ENSTR_EN, the buffer starts outputting data to start forward reasoning. During forward reasoning, the input valid signal IN_VLD and the neural network data DIN_NN are simultaneously input into the ResMLP hardware structure, where the input valid signal IN_VLD indicates the validity of the input code.

[0054] The weight RAM module Weight_RAM handles the reading and writing of weights according to the received address and enable signal.

[0055] In ResMLP-Block, the Systolic-Array based linear layer performs N×K matrix transformation. Each processing element (PE) in this layer multiplies the weights with the input and passes the data along with the partial sum to the next processing element. The matrix multiplication outputs at the end of each row are integrated through a multiplexer and output in a pipelined manner. The ReLU layer uses combinatorial logic to implement symbol determination to perform its function.

[0056] The input linear layer Input_Linear manages the input linear transformation using weight rotation and systolic array structure. Given the large input size, fully expanding the array would consume a lot of power; instead, using weight rotation and multiplier reuse techniques, the intermediate layer results are stored in neuron registers, and after a rotation cycle, the final result is output. The weights within each PE are selected from the weight Weight_RAM by the control module and systematically rotated for matrix multiplication, maintaining the systolic array structure.

[0057] The ResMLP calibration result is used as a compensation value and added to the ADC output to obtain the calibration value. The synchronization FIFO module Sync_FIFO synchronizes the ADC code to ensure alignment with the ResMLP output value, allowing the summation of the ADC code and the ResMLP output. In the data output module Data_Output_Block, the output valid signal OUT_VLD indicates the validity of the output result. The multiplexer (MUX) manages the output to ensure that the calibration value is generated only when the externally provided calibration enable CAL_EN and the ResMLP output valid OUT_VLD are both activated; otherwise, a zero value is output; the D flip-flop is used for clock synchronization.

[0058] B6. Use the trained calibration network to calibrate the unknown input signal. The spectrum result is as follows: Figure 5 As shown, it can be seen that the spurious and harmonic waves are filtered out, and the maximum spurious wave drops from -86.24dB to -101.22dB.

[0059] B7. This example collects signals with a normalized amplitude range of [0.1, 1.0] and applies the calibration technique to perform calibration. The results are as follows: Figure 6 For an input signal with an amplitude of 0.95, this method achieves a 37.14dB SFDR improvement before and after calibration; for an input signal with an amplitude of 0.13, it achieves a 21.73dB improvement.

[0060] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A global ADC calibration method based on reinforcement learning, characterized in that: A calibration network is constructed and trained using a reinforcement learning algorithm. The output signal of the ADC to be calibrated is input into the trained calibration network. The calibration network outputs a calibration value. The output signal of the ADC to be calibrated is added with the calibration value to obtain a calibrated signal.

2. The ADC global calibration method based on reinforcement learning according to claim 1, characterized in that: The calibration network processes the ADC output signal to be calibrated by: The ADC output signal to be calibrated is converted into a low-dimensional floating-point vector using a one-dimensional convolutional encoder; The low-dimensional floating-point vector is subjected to feature extraction and feature fusion through multiple cascaded three-branch modules; The signal output by the last three-branch module is mapped through a linear layer to obtain the calibration value; Among them, each three-branch module processes the output signal through the residual branch, linear branch and nonlinear branch respectively, and then fuses the processed data through weighted summation, and uses the fused value as the output of the three-branch module.

3. The ADC global calibration method based on reinforcement learning according to claim 2, characterized in that: The one-dimensional convolutional encoder consists of a cascade of one-dimensional convolutional layers and linear layers.

4. The ADC global calibration method based on reinforcement learning according to claim 2, characterized in that: The signal processing process of the three-branch module is expressed as: Y=αr×X+αl×Linear(X)+αn×ReLU(Linear(ReLU(Linear(X)))) Among them, Y represents the output of the three-branch module, X represents the input of the three-branch module; α r represents the fusion weight of the residual branch, α l represents the fusion weight of the linear branch, α n Represents the fusion weight of the nonlinear branch; Linear(·) represents the linear function; ReLU(·) represents the ReLU activation function layer.

5. The ADC global calibration method based on reinforcement learning according to claim 1 or 2, characterized in that: In the process of training the calibration network using the reinforcement learning algorithm, the ADC output signal to be calibrated is used as the state value, and the calibration amount corresponding to the ADC output signal to be calibrated is used as the action value. The calibration network is used to select the action value with the largest reward value according to the state value, and the reward value is the multi-tone spurious free dynamic range of the ADC output signal to be calibrated.

6. The ADC global calibration method based on reinforcement learning according to claim 4, characterized in that: The calculation of the multi-tone spurious free dynamic range of the ADC output signal to be calibrated includes: setting a range threshold β, taking the mean of other frequency points except the basic frequency of the spectrum diagram of the ADC output signal to be calibrated, and negating the sum of all frequency points in the other frequency points that are greater than the mean + β to obtain the multi-tone spurious free dynamic range of the ADC output signal to be calibrated.

7. An ADC global calibration circuit based on reinforcement learning, characterized in that: The circuit includes a control network and a calibration network. The control network includes an input FIFO buffer module, a neural controller module, a weight RAM module and a data output module. The calibration network calibrates the data input by the input FIFO buffer module and adds the data input by the input FIFO buffer module as the calibrated value to the data output module, wherein: An input FIFO buffer module is used to buffer the ADC output signal to be calibrated, and control the data buffered in the input FIFO buffer module to be input into the calibration network through the buffer output control signal of the neural controller module; A neural controller module is used to control the read enable, write enable, read address and write address functions of the weight RAM module, and is also used for signal control; Weight RAM module, used to store weight data needed to calibrate the network; The data output module is used to ensure that the calibrated signal output by the calibration network is output only when the calibration enable and the calibration network output are both activated externally.

8. The ADC global calibration circuit based on reinforcement learning according to claim 7, characterized in that: The calibration network processes the ADC output signal to be calibrated by: The ADC output signal to be calibrated is converted into a low-dimensional floating-point vector using a one-dimensional convolutional encoder; The low-dimensional floating-point vector is subjected to feature extraction and feature fusion through multiple cascaded three-branch modules; The signal output by the last three-branch module is mapped through a linear layer to obtain the calibration value; Among them, each three-branch module processes the output signal through the residual branch, linear branch and nonlinear branch respectively, and then fuses the processed data through weighted summation, and uses the fused value as the output of the three-branch module.

9. The ADC global calibration circuit based on reinforcement learning according to claim 8, characterized in that: The linear branch includes a linear layer, and the nonlinear branch includes a linear layer and a ReLU activation function layer. The linear layer uses a systolic array to implement N×K matrix transformation. Each processing element in the systolic array multiplies the weight by the input and passes the data together with the partial sum to the next processing element. The matrix multiplication output at the end of each row is integrated through a multiplexer and output in a pipeline manner; the ReLU activation function layer uses combinational logic to implement symbol determination.

10. The ADC global calibration circuit based on reinforcement learning according to claim 8, characterized in that: The one-dimensional convolutional encoder consists of a cascaded one-dimensional convolutional layer and a linear layer. The linear layer is implemented using weight rotation and a systolic array structure. Each processing element in the systolic array obtains the weight value from the weight RAM module. The intermediate layer results are stored in the neuron register and systematically rotated for matrix multiplication to maintain the systolic array structure. After one rotation cycle, the final result is output.

Images