EP0993671B1 - Method for searching a noise model in noisy sound signals - Google Patents

Description

The invention relates to improving the intelligibility of voice communications in the presence of noise. It no longer applies especially but not exclusively to telephone communications or by radiotelephone or other electronic means, at the speech recognition, etc., whenever the recording environment sound is noisy and may deteriorate the perception or recognition of the transmitted voice.

An example can be given about communications voices inside an airplane or other noisy vehicle. In the case from an airplane, noise comes from engines, air conditioning, ventilation of on-board equipment, aerodynamic noise. These noises are picked up by the microphone in which the pilot or a member of the crew.

The invention provides a method of searching for a model of noise which can be used in particular in treatments for reducing noise. Noise reduction treatments based on the noise model found allow to increase the signal / noise ratio of the transmitted signal, a aim being to deteriorate the intelligibility of the signal as little as possible. In this request, the denoising denoising and denoising will be used to speak operations to remove or reduce noise components present in the signal.

The denoising can be based as we will see on the permanent search for an ambient noise model, on spectral analysis of this noise, and on the digital reconstruction of a useful signal eliminating as much as possible the modeled noise.

The noise model is sought in the noisy signals themselves and whenever a plausible noise pattern has been found, this noise model is stored for use. Then a new research begins to find a more suitable model or simply more recent.

More specifically, the invention provides a search method automatic noise patterns in audio input signals noisy, in which we digitize the input signals, and we process these signals from a model found (for example in order to eliminate at best noise corresponding to the model), characterized in that the signals input are cut into successive frames of P samples each, and a repetitive search for a noise model is performed in permanence in the input signals themselves, looking for N successive frames having the expected characteristics of a noise, in storing the corresponding NxP samples to constitute a model of noise useful for processing denoising of input signals, and repeating the research to find a new noise model and store the new one model to replace the previous one or keep the previous model according to the respective characteristics of the two models.

Consequently, the noise model used in particular for denoising is not a known predetermined model or a chosen model among several predetermined models, but this is a model found in the noisy signal itself, which allows not only to adapt denoising to the real annoying noise, but also to adapt the denoising to variations of this noise.

The noise model is obtained by considering that the signals whose energy is stable (and preferably, as we will see, whose energy is minimal), over a certain period probably represent noise; the invention is characterized, according to claim 1, in that the searching for a noise model then includes searching for N frames successive whose energies are close to each other (N being between a minimum value N1 and a maximum value N2), the calculation of the average energy of the N successive frames found, and the storage NxP samples as a new active model if the relationship between this average energy and the average energy of the frames of the active model previously stored is below a determined replacement threshold.

Note that documents US-A-5,550,924 and WO 97 18,647 use spectral analysis to search for noise patterns.

The search for N successive frames then comprises at minus the following iterative steps: calculation of the energy of a frame current of rank n can be added to a current model of preparation already comprising n-1 successive frames; ratio calculation between this energy and the energy of the previous frame of rank n-1 (and of preferably that of other previous frames between 1 and n-1); comparison of this ratio with a low threshold less than 1 and a high threshold greater than 1; and decision on the possibility of incorporating the frame of rank n in the model in being developed: the frame is not incorporated into the model if the report is not between the two thresholds; it is incorporated into the model if the ratio is between the two thresholds. The procedure is repeated on the next current frame of input signals, with incrementation of n, until the model is stopped.

The development of the model is stopped either in the case where n reaches the high value N2, either in the case where the frame of rank n is not incorporated into the model because the calculated energy ratio goes out of prescribed range. In the latter case, the model developed cannot be taken count as an active model only if n-1 is already greater than or equal to the minimum N1, because the principle is that a noise model is representative if it has an approximately stable energy on at least N1 frames.

Preferably, the model developed does not become active in place of the previous model only if the ratio between its average energy per frame and the average energy of the previous model does not exceed a threshold of predetermined replacement.

In any case, the search for a new model starts again as soon as the preparation of the previous one is interrupted.

Finally, preferably, it can be provided that the replacement of a previous model by a new model is inhibited as soon as speech is detected in noisy signals. The presence of speech can indeed be detected by digital signal processing procedures (such as than those that can be used in speech recognition).

• la figure 1 représente un organigramme général d'un procédé de réduction de bruit utilisant le procédé de l'invention;
• la figure 2 représente un exemple typique de signal issu d'une prise de son bruitée;
• la figure 3 représente l'organigramme des étapes de recherche d'un modèle de bruit dans le signal d'entrée;
• la figure 4 représente un exemple d'architecture de circuit électronique pour la mise en oeuvre d'opérations de débruitage utilisant le procédé selon l'invention.

Other characteristics and advantages of the invention will appear on reading the detailed description which follows and which is given with reference to the appended drawings in which:

• FIG. 1 represents a general flow diagram of a noise reduction method using the method of the invention;
• FIG. 2 represents a typical example of a signal coming from a noisy sound recording;
• FIG. 3 represents the flow diagram of the steps for searching for a noise model in the input signal;
• FIG. 4 represents an example of electronic circuit architecture for the implementation of denoising operations using the method according to the invention.

In speech analysis, it is usual to consider that the stationary sound production systems are established over time between 10 and 20 milliseconds.

The signal analysis which allows denoising will be based on spectral analysis of signals in time intervals of duration D, which we will call "frames", and which will have approximately this duration.

Each frame will comprise P = 2 ^p samples of digitized signal, the number P depending on the sampling frequency of the processed signal, so that the frame has a duration of the order of 10 to 20 ms regardless of the frequency of sampling F _e = 1 / T _e . For example, for a sampling frequency of 10 kHz, the frame will contain P = 128 samples (p = 7) and will last 12.8 ms.

The diagram in Figure 1 is a flowchart explaining the general principle of the denoising process.

The input signal to be processed, coming for example from a microphone, is noted u (t), with a useful part s (t) and an undesirable noise b (t), with u (t) = s (t) + b (t), time t being assumed to be discrete (t = kT _e ) since the signal is sampled before being digitized in an analog-digital converter.

In the following, we will consider, as an example representing the main application of the invention, that the processing of input signals is a denoising treatment based on the noise model found. other applications can be considered (search for whistling consonants or hushing, for example).

The general principle of the denoising process is based on a permanent and automatic search for a noise model which will be used to process the input signal to denois it. This research is done on digitized u (t) signal samples stored in a buffer input. This memory is capable of memorizing all the samples of several frames of the input signal (e.g. at least 2 frames).

The noise model sought consists of a succession multiple frames including energy stability and energy level relative suggest that it is an ambient noise and not a speech signal or some other disturbing noise. We will see later how this automatic search.

When a noise pattern is found, all samples from the N successive frames representing this noise model are kept in memory, so the spectrum of this noise can be analyzed and can be used to denoising. But the automatic noise search continues from the input signal u (t) to possibly find a newer model and more suitable, either because it better represents ambient noise, or because that ambient noise has changed. The newer noise model is implemented memory in place of the previous one, if the comparison with the previous one shows that it is more representative of ambient noise.

The denoising of the input signal u (t) is done from the model of noise that is in memory, and more precisely from the characteristics spectral of this model. A Fourier transform and an estimate of average spectral density of noise are therefore performed on the model of stored noise. The denoising operation is preferably done using a digital filtering from Wiener which will be discussed in more detail. The filter of Wiener is parameterized by the spectral characteristics of the model of noise recorded and by the spectral characteristics of the signal u (t) to denoise. The digitized input signal therefore undergoes a transform of Fourier and an estimate of spectral density. The numerical values of the Fourier transform, i.e. the input signal represented by its frequency components, are processed by the Wiener filter and the output of the Wiener filter represents, in frequency space, the signal digital denoised, that is to say rid as much as possible of the noise represented by the registered model.

The filtered digital signal is used either for the reconstruction of a sound signal in which the ambient noise has been partly eliminated, i.e. at the speech Recognition.

The phase of automatic search for a noise model and the permanent updating of this model are crucial steps in the process and are more precisely the subject of the invention.

• le bruit qu'on veut éliminer est le bruit de fond ambiant;
• le bruit ambiant a une énergie relativement stable à court terme,
• la parole est le plus souvent précédée d'un bruit de respiration du pilote qu'il ne faut pas confondre avec le bruit ambiant; mais ce bruit de respiration s'éteint quelques centaines de millisecondes avant la première émission de parole proprement dite, de sorte qu'on ne retrouve que le bruit ambiant juste avant l'émission de parole;
• et enfin, les bruits et la parole se superposent en termes d'énergie de signal, de sorte qu'un signal contenant de la parole ou un bruit perturbateur, y compris la respiration dans le microphone, contient forcément plus d'énergie qu'un signal de bruit ambiant.

The starting postulates for the automatic development of a noise model are as follows:

• the noise that we want to eliminate is the ambient background noise;
• ambient noise has a relatively stable energy in the short term,
• speech is most often preceded by a breathing noise from the pilot which should not be confused with ambient noise; but this breathing noise is extinguished a few hundred milliseconds before the first speech emission proper, so that only the ambient noise is found just before the speech emission;
• and finally, noise and speech are superimposed in terms of signal energy, so that a signal containing speech or disturbing noise, including breathing in the microphone, necessarily contains more energy than a ambient noise signal.

As a result, we will make the following simple hypothesis: noise ambient is a signal with a stable minimum energy in the short term. By short term we need to understand some frames, and we will see in the example practice given below that the number of frames intended to assess the noise stability is 5 to 20. Energy must be stable over several frames, otherwise we must assume that the signal contains rather speech or noise other than ambient noise. It must be minimal, fault what we consider that the signal contains breathing or phonetic speech elements resembling noise but overlapping to ambient noise.

Figure 2 shows a typical evolution configuration temporal energy of a microphone signal at the time of a start speech emission, with a breath noise phase, which goes out for a few tens to hundreds of milliseconds to make room for the ambient noise alone, after which a high energy level indicates the presence speech, to finally return to ambient noise.

The automatic search for ambient noise then consists of find at least N1 successive frames (for example N1 = 5) whose energies are close to each other, that is to say that the relationship between the signal energy contained in a frame and the signal energy contained in the or, preferably, the previous frames is located inside of a determined range of values (for example between 1/3 and 3). When such a succession of relatively stable energy frames has been found, we store the numerical values of all the samples of these N frames. This set of NxP samples constitutes the current model of noise. It is used in denoising. Analysis of the following frames keep on going. If we find another succession of at least N1 frames successive ones meeting the same energy stability conditions (weft energy ratios in a given range), we compare then the average energy of this new succession of frames to the energy average of the stored model, and we replace the latter with the new succession if the ratio between the average energy of the new succession and the average energy of the stored model is below a threshold of determined replacement which can be 1.5 for example.

From this replacement of a noise model by a more recent less energetic or not much more energetic, it follows that the noise model is generally based on permanent ambient noise. Even before speaking, preceded by breathing, there is a phase where ambient noise alone is present for a sufficient time to be taken into account as an active noise model. This phase ambient noise alone after breathing is brief; the number N1 is chosen relatively weak, so that we have time to readjust the noise model on ambient noise after the breathing phase.

If the ambient noise changes slowly, the change will be taken into account. account of the fact that the comparison threshold with the stored model is greater than 1. If it evolves more rapidly in the increasing direction, the evolution may not be taken into account, so it is best to plan to reset the search for a model from time to time noise. For example, in a stopped ground plane, the ambient noise will be relatively weak, and it should not be that during the phase of takeoff the noise model remains frozen on what it was at a standstill because a noise model is only replaced by a less energetic model or not much more energetic. The methods of reset envisaged.

FIG. 3 represents a flowchart of the operations of automatic search for an ambient noise pattern.

The input signal u (t), sampled at the frequency F _e = 1 / T _e and digitized by an analog-digital converter, is stored in a buffer memory capable of storing all the samples of at least 2 frames.

The number of the current frame in an operation of looking for a noise pattern is denoted by n and is counted by a counter as you search. At the initialization of the search, n is set to 1. This number n will be incremented progressively the development of a model of several successive frames. when analyzes the current frame n, the model already understands by hypothesis n-1 successive frames meeting the conditions imposed to be part of a model.

We first consider that this is a first elaboration of model, no other previous model having been built. We'll see then what happens for further elaborations.

The signal energy of the frame is calculated by summing the squares of the numerical values of the samples of the frame. She is kept in memory.

We then read the next frame of rank n = 2, and its energy is calculated in the same way. It is also kept in memory.

The ratio between the energies of the two frames is calculated. If this ratio is between two thresholds S and S 'one of which is greater than 1 and the other is less than 1, we consider that the energies of the two frames are close and that the two frames can be part of a noise model. The thresholds S and S 'are preferably inverse to each other (S' = 1 / S) by so you just have to define one to get the other. For example, a value typical is S = 3, S '= 1/3. If the frames can be part of the same noise model, the samples that compose them are stored for start building the model, and the research continues iterating through incrementing n by one.

If the relationship between the energies of the first two frames goes out of the imposed interval, the frames are declared incompatible and the search is reset by resetting n to 1.

In the case where the search continues, the rank n is incremented of the current frame, and we perform, in a procedure loop iterative, an energy calculation of the next frame and a comparison with the energy of the previous frame or previous frames, using the thresholds S and S '.

Note in this connection that two types of comparison are possible to add a frame to n-1 previous frames which have already been considered homogeneous in energy: the first type of comparison consists in comparing only the energy of the frame n to the energy of the n-1 frame. The second type is to compare the energy of frame n at each of frames 1 to n-1. The second way leads to greater homogeneity of the model but it has the disadvantage of not not take sufficiently into account the cases where the noise level increases or decreases rapidly.

Thus, the energy of the frame of rank n is compared with the energy of the frame of rank n-1 and possibly of other frames previous (not necessarily all of them for that matter).

• ou bien n est inférieur ou égal à un nombre minimal N1 en dessous duquel le modèle ne peut pas être considéré comme significatif du bruit ambiant parce que la durée d'homogénéité est trop courte; par exemple N1 = 5; dans ce cas on abandonne le modèle en cours d'élaboration, et on réinitialise la recherche au début en remettant n à 1;
• ou bien n est supérieur au nombre minimal N1. Dans ce cas, puisqu'on trouve maintenant un manque d'homogénéité, on considère qu'il y a peut-être un début de parole après une phase de bruit homogène, et on conserve à titre de modèle de bruit tous les échantillons des n-1 trames de bruit homogènes qui ont précédé le manque d'homogénéité. Ce modèle reste stocké jusqu'à ce qu'on trouve un modèle plus récent qui semble également représenter du bruit ambiant. La recherche est réinitialisée de toutes façons en remettant n à 1.

If the comparison indicates that there is no homogeneity with the previous frames, because the energy ratio is not between 1 / S and S, two cases are possible:

• or n is less than or equal to a minimum number N1 below which the model cannot be considered as significant of the ambient noise because the duration of homogeneity is too short; for example N1 = 5; in this case we abandon the model being developed, and we reset the search at the beginning by giving n to 1;
• or n is greater than the minimum number N1. In this case, since we now find a lack of homogeneity, we consider that there may be a start of speech after a homogeneous noise phase, and we keep as a noise model all the samples of the n -1 homogeneous noise frames which preceded the lack of homogeneity. This model remains stored until a newer model is found which also appears to represent ambient noise. The search is reset anyway by resetting n to 1.

But the comparison of frame n with the previous ones would have could still lead to the observation of a still homogeneous frame in energy with the previous one (s). In this case, either n is less than a second number N2 (for example N2 = 20) which represents the length desired noise model, or n has become equal to this number N2. The number N2 is chosen so as to limit the calculation time in subsequent noise spectral density estimation operations.

If n is less than N2, the homogeneous frame is added to the to help build the noise model, n is incremented and the next frame is analyzed.

If n is equal to N2, the frame is also added to the n-1 previous homogeneous frames and the model of n homogeneous frames is stored for use in noise elimination. Searching for a model is also reset by resetting n to 1.

The previous steps relate to the first search for model. But once a model has been stored, it can at any time be replaced by a more recent model.

The replacement condition is still a condition of energy, but this time it relates to the average energy of the model and not more about the energy of each frame.

Therefore, if a possible model has just been found, with N frames where N1 <N <N2, we calculate the average energy of this model which is the sum of the energies of the N frames, divided by N, and we compare it to the average energy of the N 'frames of the previously stored model.

If the ratio between the average energy of the new model possible and the average energy of the current model in use is less than one SR replacement threshold, the new model is considered better and we store it in place of the previous one. Otherwise, the new model is rejected and the old remains in force.

The threshold SR is preferably slightly greater than 1.

If the SR threshold was less than or equal to 1, we would store at each times the least energetic homogeneous frames, which corresponds well to the fact that ambient noise is considered to be the energy level at below which we never descend. But, we would eliminate any possibility evolution of the model if the ambient noise starts to increase.

If the SR threshold was too high above 1, there is a risk of poorly distinguish ambient noise and other disturbing noises (breathing), or even some phonemes that sound like noise (consonants hissing or hissing for example). Noise removal from a noise pattern stalled on breath or on whistling consonants or hissing could then harm the intelligibility of the denoised signal.

In a preferred example the threshold SR is approximately 1.5. At above this threshold we will keep the old model; below this threshold we will replace the old model with the new one. In both cases, we will reset the search by restarting the reading of a first frame of the input signal u (t), and by putting n at 1.

To make the development of the noise model more reliable, we can provide that the search for a model is inhibited if an emission of speech is detected in the wanted signal. The digital treatments of commonly used signal in speech detection identify the presence of words based on the characteristic spectra of periodicity of certain phonemes, in particular the corresponding phonemes to vowels or voiced consonants.

The purpose of this inhibition is to prevent certain sounds from being taken for noise when these are useful phonemes, that a model of noise based on these sounds is stored and that noise suppression after the development of the model then tends to suppress all the sounds Similar.

In addition, it is desirable to plan from time to time a reset of the model search to allow a reset model day when ambient noise increases were not taken into account that SR is not much greater than 1.

Ambient noise can indeed increase significantly and fast, for example during the acceleration phase of the engines of a plane or other vehicle, air, land or sea. But the SR threshold requires that the previous noise model be kept when the energy average noise increases too quickly.

If we wish to remedy this situation, we can proceed with different ways but the simplest way is to reset the model periodically by researching and imposing a new model as an active model regardless of the comparison between this model and the previously stored model. Periodicity can be based on duration average speech in the intended application; for example the durations of speech are on average a few seconds for the crew of a airplane, and the reset can take place with a periodicity of a few seconds.

The proper denoising treatment, carried out from of a stored noise model, can be performed as follows, by working on the Fourier transforms of the input signal.

The Fourier transform of the input signal is carried out frame by frame and provides for each frame P samples in the frequency space, each sample corresponding to a frequency F _e / i with i varying from 1 to P. These P samples will be processed preferably in a Wiener filter. The Wiener filter is a digital filter of P coefficients each corresponding to one of the frequencies F _e / i of the frequency space. Each sample of the input signal in the frequency space is multiplied by the respective coefficient W _i of the filter. The set of P samples thus processed constitutes a denoised signal frame, in the frequency space. For speech recognition applications, these denoised frames are used directly in the frequency space. For applications where we want to reconstruct a real denoised sound signal, we perform successively an inverse Fourier transform on each frame, a digital-analog conversion, and a smoothing.

The coefficients W _i of the Wiener filter are calculated from the spectral density of the noisy input signal and the noise spectral density of the stored noise model.

The spectral density of a frame of the input signal is obtained from the Fourier transform of the noisy input signal. For each frequency, we take the squared module of the sample provided by the Fourier transform, to obtain a value DS _i for each frequency F _e / i.

For the spectral density of the noise model, the module squared of the P samples is calculated for each frame, and the N squared modules corresponding to the same frequency F _e / i are averaged over the N frames of the noise model. P noise density values DB _{i are} obtained.

The Wiener coefficient W _i for the frequency F _e / i is then W ₁ = 1 - DB _i / DS _i .

The sample of rank i of the Fourier transform of an input signal frame is multiplied by W _i and the succession of the P samples thus multiplied by P Wiener coefficients constitutes the denoised input frame.

The implementation of the method according to the invention can be done at from non-specialized computers, provided with calculation programs required and receiving the digital signal samples as they are supplied by an analog-to-digital converter.

This implementation can also be done from a specialized computer based on digital signal processors, which allows more signals to be processed more quickly digital.

FIG. 4 represents an example of general architecture of a specialized computer receiving the sound signal to denois and providing real time an audible noise signal.

The computer includes two signal processors digital DSP1 and DSP2 and working memories associated with these processors.

Noise signals are passed through a converter analog-digital CA / D and are stored in parallel in two FIFO1 and FIFO2 buffers (of the "first-in, first-out" type, i.e. first in first out). One of the memories is connected to the processor DSP1, the other to the DSP2 processor.

The DSP1 processor is the master processor and it is dedicated rather looking for a noise model. It is therefore programmed to execute at least the following operations: frame energy calculation, energy averaging, comparison with thresholds, comparison frame rank with N1 and N2, etc. It also calculates densities energy spectral of the noise model. This DSP1 processor is coupled to a dynamic working memory DRAM1 in which we store the current frame sample during a calculation, the energy of a frame current, the energy of the previous frame (s), the samples of Fourier transform of the noise model. It is also coupled with a static working memory in which the tables used are stored the computation of Fourier transforms, and the comparison thresholds S and SR.

The DSP2 processor is dedicated rather to the calculation of transforms Fourier signal to denois, calculating the spectral density of this signal, calculating Wiener coefficients, Wiener filtering, and inverse Fourier transform if the latter is to be performed. The DSP2 processor is coupled to a dynamic working memory DRAM2 and a static working memory SRAM2. DRAM2 memory stores current frame samples, transform calculation results from Fourier, results of calculation of spectral energy density of the signal, the calculated Wiener coefficients, etc. The SRAM2 memory stores in particular tables used for the computation of Fourier transforms.

The denoised sound signal samples calculated by the DSP2 processor are transmitted, through a circulating buffer FIFO3, to a digital analog converter CNIA, and to a circuit of smoothing which reconstructs the denoised sound signal in analog form.

Claims (8)

1. Process for automatically searching for noise models in noisy audio input signals, comprising the digitizing of the input signals, and the processing of these signals on the basis of a model found, in which process the input signals are chopped into successive frames of P samples each, and a repetitive search for a noise model is performed continuously in the input signals themselves, by searching for N successive frames having the expected characteristics of a noise, by storing the N×P corresponding samples so as to construct a noise model useful in the denoising processing of the input signals and by iteratively repeating the search so as to find a new noise model and store the new model as replacement for the previous one or retain the previous model according to the respective characteristics of the two models, characterized in that the search for a noise model comprises the search for N successive frames whose energies are close to one another, N lying between a minimum value N1 and a maximum value N2, the calculation of the average energy of the N successive frames found, and the storing of the N×P samples in the guise of new active model if the ratio between this average energy and the average energy of the frames of the active model previously stored is less than a determined replacement threshold.
2. Process according to Claim 1, characterized in that the search for N successive frames then comprises at least the following iterative steps: calculation of the energy of a current frame of rank n able to be appended to a model undergoing formulation already comprising n-1 successive frames; calculation of the ratio between this energy and the energy of the previous frame of rank n-1; comparison of this ratio with a low threshold less than 1 and a high threshold greater than 1; and decision regarding the possibility of incorporating the frame of rank n into the model undergoing formulation as a function of the result of the comparison.
3. Process according to Claim 2, characterized in that the search for N successive frames also comprises the calculation of the ratio between the energy of the current frame and the energy of one or more other previous frames, the comparison with the thresholds, the frame being incorporated into the model undergoing formulation as a function of the result of the comparison.
4. Process according to one of Claims 2 and 3, characterized in that in the case where the frame of rank n is incorporated into the model, n is incremented by one unit so as to continue the formulation of the model if n is less than N2, and, in the contrary case, the formulation of the model is halted, the average energy of the n frames is calculated, the ratio between this energy and the average energy of the frames of the previously stored model is calculated, the previous model is retained or is replaced by the model undergoing formulation according to the value of the ratio, and the iterative search for a new model is restarted.
5. Process according to one of Claims 2 and 3, characterized in that in the case where the current frame of rank n is not incorporated into the model undergoing formulation,

   the formulation of the model of n-1 frames is halted;

   if n is greater than N1, the ratio between the average energy of the frames of the model undergoing formulation and the average energy of the frames of the previously stored model is calculated, and the previous model is. retained or is replaced by the new model according to the value of the ratio,

   and an iterative search for a new model is restarted.
6. Process according to one of the preceding claims, characterized in that a search is made for the presence of speech in the input signal, and the search for a new model is disabled if the presence of speech is detected.
7. Process according to one of the preceding claims, characterized in that the search is periodically reinitialized by imposing the new model regardless of the respective characteristics of the new model and of the previous model.
8. Process according to one of the preceding claims, characterized in that the noisy input signals are processed on the basis of a found noise model, by spectral filtering, with a view to eliminating as far as possible the noise corresponding to the model.

Images