High-speed target tracking method based on pulse sequence image sensor

Abstract

The present invention relates to the fields of optics, image sensor imaging and image processing, and proposes a high-speed target tracking method for a pulse image sensor, which can effectively improve the accuracy and stability of the pulse image sensor in tracking high-speed targets in a scene. To this end, the technical solution adopted by the present invention is a high-speed target tracking method based on a pulse sequence image sensor, and the steps are as follows: (1) obtaining the pulse interval and pulse frequency according to the pulse data; (2) performing motion detection on the moving target, and removing ghosts and holes that appear during the motion detection process; (3) locking and tracking the target; (4) after finding the real target area, using bilateral filtering to reconstruct the image of the scene and the target. The present invention is mainly used in image sensor design and manufacturing occasions.

Description

Technical Field

[0001] The present invention relates to the fields of optics, image sensor imaging and image processing, and in particular to a high-speed target tracking method based on a pulse sequence image sensor. Background Art

[0002] Image sensors have always been a hot topic in human research. As a neuromorphic visual sensor, the bionic pulse train image sensor has the characteristics of high frame rate and low data throughput, which meets the needs of high-speed imaging. The circuit structure of the pulse train image sensor pixel, its equivalent model and working principle are as follows: Figure 1 As shown in Figure 1, the circuit structure of the pixel mainly includes a photodiode, a reset tube, a comparator, a self-reset unit, and a pixel readout circuit unit. It can be seen from the circuit equivalent model and working principle that under illumination conditions, the integrator continuously integrates the photocurrent I D , generating photogenerated charge Q D , when the photogenerated charge reaches the threshold Q ref When the integrator is reset and restarts integration, a pulse data 1 is generated and transmitted to the outside of the chip when the synchronous readout signal arrives. Since the time period for the arrival of the readout signal is one frame period, if no pulse data 1 is generated within one frame period, the pulse data 0 is output to the outside of the chip when the readout signal arrives. Therefore, the output data of the pulse image sensor is a single-bit data sequence of only 0 and 1.

[0003] Imaging system and analysis algorithm are important components of machine vision, and the two are interdependent and inseparable. High-speed target tracking system is an important branch of machine vision. In high-speed target tracking system, image sensor is used as imaging system, and target tracking algorithm is used as analysis algorithm. In the past few decades, although target tracking algorithm has been extensively studied as analysis algorithm and has achieved satisfactory reliability and accuracy, most of the literature is based on the study of video image sequences taken by traditional cameras, while there are very few studies on bionic pulse image sensors. On the one hand, the data of pulse image sensor is single-bit data and does not directly contain grayscale information, so it is not suitable for traditional target tracking algorithm. On the other hand, in order to apply traditional target tracking method, it is necessary to reconstruct the image sequence based on single-bit data, although some studies have proposed some related image reconstruction algorithms based on pulse image sensor, which can reconstruct the grayscale information of the scene based on single-bit pulse data. However, if the traditional target tracking algorithm is applied to the reconstructed image sequence, the error and noise introduced in the image reconstruction process will seriously affect the accuracy and stability of target tracking. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to propose a high-speed target tracking method for a pulse image sensor, which can effectively improve the accuracy and stability of the pulse image sensor in tracking high-speed targets in a scene. To this end, the technical solution adopted by the present invention is a high-speed target tracking method based on a pulse sequence image sensor, the steps are as follows:

[0005] (1) Obtaining pulse interval and pulse frequency according to pulse data;

[0006] (2) Perform motion detection on moving targets and remove ghosts and holes that appear during the motion detection process;

[0007] (3) Locking and tracking the target;

[0008] (4) After finding the real target area, bilateral filtering is used to reconstruct the image of the scene and the target.

[0009] The detailed steps are as follows:

[0010] (1) Obtain the pulse interval and pulse frequency based on the pulse data. For the pixel (i, j) at time t, first find out the output time Zmax of the most recent pulse data 1 before time t and the output time Zmin of the most recent pulse data 1 after time t, as shown in formula (1):

[0011]

[0012] Among them I z (i, j) is the pulse data of pixel (i, j) at time z, and then the original pulse interval P of the pixel at time t t (i, j) is obtained by the difference between these two moments, as shown in formula (2):

[0013] P t (i,j)=Zmin-Zmax (2)

[0014] The pulse data is converted into pulse intervals by formula (2), and then the pulse frequency is calculated according to the converted pulse intervals, that is:

[0015] f t (i,j)=1 / P t (i,j) (3)

[0016] (2) Detect moving targets by using the pulse frequency as the pixel value and use the pixel value of frame A to initialize the background model, as shown in formula (4):

[0017] M A (i,j)={f1(i,j),f2(i,j),…,f A(i,j)} (4)

[0018] Then calculate the Euclidean distance between the pixel value at the current moment and the background model, and compare the calculated result with the set threshold R D Compared with the threshold R D , then the comparison result r k is 1, which is higher than the threshold R D If the statistical result is greater than the given threshold R, then B , then the detection result of the pixel is B t (i, j) is the background, and the binary grayscale value is 0. Otherwise, it is a moving target, and the binary grayscale value is 255, that is:

[0019]

[0020] At the same time, if the pixel is detected as background, the background model needs to be updated, that is, the pixel value is used to randomly replace any sample in the background model of any pixel in the neighborhood of the pixel, as shown in formula (6):

[0021]

[0022] where f an (i,j) is the background model M of pixel (i,j) A For any sample in (i,j), f an (i an ,j an ) is any sample in the background model of any pixel in the neighborhood of pixel (i, j);

[0023] Then remove the ghosts and holes that appear in the motion detection process. For pixel (i, j), the neighborhood R of pixel (i, j) is ij Fixed time period t h The pulse signal within the neighborhood is counted and the pulse density d of the neighborhood is calculated. t (i,j):

[0024]

[0025] Among them, H t (x,y) is the neighborhood R ij The pulse data of the inner pixel (x, y), the pulse density range at the actual moving target is R H , then although the binary grayscale value of the ghost is 255, its pulse density must not be within R H Within the range, ghosting is eliminated accordingly, namely:

[0026]

[0027] Similarly, since the binary grayscale value of the hole inside the target is 0, its pulse density is the same as that of the actual target, the hole is compensated accordingly, that is:

[0028] B t (i,j)=255,if B t (i,j)=0 and d t (i,j)∈R H (9)

[0029] (3) Locking and tracking the target: First, lock the target to be tracked from all the moving objects detected by motion, and mark the area where the target is located. Then, use the target area at this time as the target model and calculate the probability density of the pulse interval of the area. After one or several frames, mark the areas where all the moving objects detected by motion are located again, and then calculate the probability density of the pulse interval of each moving object area and compare it with the probability density of the target model. Then, the moving object with the most similar probability density to the target model within a certain distance range is the target to be tracked; finally, use the moving object as the new target model, and use the probability density in the moving object area as the probability density of the target model to update the target model for target tracking in the next frame or several frames;

[0030] (4) After finding the real target area, bilateral filtering is used to reconstruct the scene and target image, as shown in formula (17):

[0031]

[0032] Where P′ t (i,j) The pulse interval after filtering, P t (i, j) is the original pulse interval, W p is the normalization coefficient, G σs is the spatial kernel, G σr is the pixel kernel, S is the spatial domain of pixel (i, j), σs and σr are the standard deviations of the spatial kernel and the pixel kernel, and then the image is reconstructed as shown in formula (18):

[0033] G t (i,j)=K / P′ t (i,j),K=255×P min (18)

[0034] Where K is the conversion coefficient from pulse interval to grayscale value, P min is the minimum pulse interval in the pulse interval sequence;

[0035] In step (3), the target to be tracked is first locked at the initial time t0, the area where the target is located is marked by the connected domain, and the target area at this time is used as the target model and named R Q , the nth pixel in the area is named p n , whose pulse interval is z n , the center pixel of the area is named p0, and its pulse interval is z0, as shown in formula (10):

[0036]

[0037] Then calculate the pulse interval probability density Q of the target model u (p0), as shown below:

[0038]

[0039] Where K is the Epannechnikov kernel function, h is the window size of the kernel function, and C is the normalization coefficient, so that b(z n ) is the pulse interval z of the pixels in the area n The histogram quantization value, Nbin is the maximum quantization range, u is the eigenvalue, δ(b(z n )-u) is the Delta function, which is used to determine the quantized value b(z n ) belongs to eigenvalue u, N um is the number of all pixels in the area. After one or several frames, that is, at time t, the area where all moving objects detected by motion are located is marked again through the connected domain and named R P1 , R P2 , R P3 , …, R Pm ,…, assuming that the mth moving object region R Pm The nth pixel in is p nm , whose pulse interval is z nm , the center pixel of the model is p 0m , whose pulse interval is z 0m , as shown in formula (12):

[0040]

[0041] Then the pulse interval probability density of all moving object areas is calculated, where the probability density Q of the mth moving object is um (p 0m ) is shown in formula (13):

[0042]

[0043] The similarity function Bhattacharyya coefficient is used to measure the target model and all moving objects R P1 , R P2 , R P3 , …, R Pm , ..., and find the real target R P In addition, since the pulse image sensor has a high frame rate, the position change of the target between two or several frames must be small, that is, the target between two or several frames must remain within a certain distance range L d Therefore, the Bhattacharyya coefficient can be adjusted as follows:

[0044]

[0045] where ρ m is the adjusted Bhattacharyya coefficient value of the mth candidate model. In addition, L d It can be adjusted adaptively and iteratively. First, L d Set a very small initial value, and then according to ρ m The value of L is adjusted d If the value of m The maximum value of is 0, then L d The size increases by 1 until ρ m The maximum value of is not 0, as shown in formula (15):

[0046]

[0047] The real target R P Then ρ m The largest moving object is:

[0048] R P =R Pm ,if ρ m =max(ρ m ),m∈[1,2,3,…] (16).

[0049] Then update the target model, i.e. R Q =R P .

[0050] The characteristics and beneficial effects of the present invention are:

[0051] A high-speed target tracking method is proposed for pulse sequence image sensors. This method can effectively improve the accuracy and stability of pulse image sensors in tracking high-speed targets in the scene. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1Circuit structure, equivalent model and working principle of image sensor pixels.

[0053] Figure 2 The process of converting pulse data to pulse interval.

[0054] Figure 3 Lock and mark the target.

[0055] Figure 4 Mark all moving targets.

[0056] Figure 5 The overall process of the algorithm.

[0057] Figure 6 Motion detection of high-speed turntables.

[0058] (a) Shooting scene of a high-speed turntable (b) Before ghost and hole removal (c) After ghost and hole removal.

[0059] Figure 7 The tracking effect of the algorithm on a high-speed turntable. DETAILED DESCRIPTION

[0060] The technical solution of the present invention is as follows:

[0061] (1) First, we need to obtain the pulse interval and pulse frequency based on the pulse data. Taking the pixel (i, j) at time t as an example, we first need to find the output time Zmax of the most recent pulse data 1 before time t and the output time Zmin of the most recent pulse data 1 after time t, as shown below:

[0062]

[0063] Among them I z (i, j) is the pulse data of pixel (i, j) at time z. Then the original pulse interval P of the pixel at time t is t (i,j) can be obtained by the difference between these two moments as follows:

[0064] P t (i,j)=Zmin-Zmax (2)

[0065] The pulse data can be converted into pulse intervals by equation (2): Figure 2 Describe the conversion process of pulse data to pulse interval. Then, according to the converted pulse interval, the pulse frequency can be calculated, that is:

[0066] f t (i,j)=1 / P t (i,j) (3)

[0067] (2) Motion detection of moving targets. Since the greater the light intensity, the greater the pulse frequency, the pulse frequency can be used as the pixel value to detect moving targets. That is, the value of the pixel in frame A is used to initialize the background model, as shown below:

[0068] M A (i,j)={f1(i,j),f2(i,j),…,f A (i,j)} (4)

[0069] Then calculate the Euclidean distance between the pixel value at the current moment and the background model, and compare the calculated result with the set threshold R D Compared with the threshold R D , then the comparison result r k is 1, which is higher than the threshold R D Then the statistical comparison results are counted and if the statistical result is greater than the given threshold R B , then the detection result of the pixel is B t (i, j) is the background, and the binary grayscale value is 0. Otherwise, it is a moving target, and the binary grayscale value is 255, that is:

[0070]

[0071] At the same time, if the pixel is detected as background, the background model needs to be updated, that is, the pixel value is used to randomly replace any sample in the background model of any pixel in the neighborhood of the pixel, as shown below:

[0072]

[0073] where f an (i,j) is the background model M of pixel (i,j) A For any sample in (i,j), f an (i an ,j an ) is any sample in the background model of any pixel in the neighborhood of pixel (i,j).

[0074] Then remove the ghosts and holes that appear during motion detection. Taking pixel (i, j) as an example, we calculate the neighborhood R of pixel (i, j) ij Fixed time period t h The pulse signal within the neighborhood is counted and the pulse density d of the neighborhood is calculated. t (i,j):

[0075]

[0076] Among them, H t (x,y) is the neighborhood R ijThe pulse data of the inner pixel (x, y) is different from the light intensity at the ghost and the actual moving target, so their pulse density is also different. Assume that the pulse density range at the actual moving target is R H , then although the binary grayscale value of the ghost is 255, its pulse density must not be within R H The ghost image can be eliminated based on this, namely:

[0077]

[0078] Similarly, since the binary grayscale value of the hole inside the target is 0, its pulse density is the same as that of the actual target, the hole can be compensated accordingly, that is:

[0079] B t (i,j)=255,if B t (i,j)=0 and d t (i,j)∈R H (9)

[0080] This completes the motion detection.

[0081] (3) Lock and track the target. First, lock the target to be tracked from all the moving objects detected by motion, and mark the area where the target is located. Then, use the target area at this time as the target model, and calculate the probability density of the pulse interval of the area. After one or several frames, mark the areas where all the moving objects detected by motion are located again, and then calculate the probability density of the pulse interval of each moving object area, and compare it with the probability density of the target model. Then, the moving object with the most similar probability density to the target model within a certain distance range is the target to be tracked. Finally, use the moving object as the new target model, and use the probability density in the moving object area as the probability density of the target model to update the target model for target tracking in the next frame or several frames. Take the tracking of the turntable pattern as an example, if Figure 3 As shown in the figure, it is assumed that the aircraft pattern on the turntable is locked as the target to be tracked at the initial time t0, and the area where the target is located is marked by the connected domain, and the target area at this time is used as the target model and named R Q , the nth pixel in the area is named p n , whose pulse interval is z n , the center pixel of the area is named p0, and its pulse interval is z0, as shown below:

[0082]

[0083] Then calculate the pulse interval probability density Q of the target model u (p0), as shown below:

[0084]

[0085] Where K is the Epannechnikov kernel function, h is the window size of the kernel function, and C is the normalization coefficient, so that b(z n ) is the pulse interval z of the pixels in the area n The histogram quantization value, Nbin is the maximum quantization range, u is the eigenvalue, δ(b(z n )-u) is the Delta function, which is used to determine the quantized value b(z n ) belongs to eigenvalue u, N um is the number of all pixels in the area. After one or several frames, Figure 4 As shown, at time t, the area where all the motion patterns on the turntable are located is detected again through the connected domain marking motion, and it is named R P1 , R P2 , R P3 , …, R Pm ,…, assuming that the mth moving object region R Pm The nth pixel in is p nm , whose pulse interval is z nm , the center pixel of the model is p 0m , whose pulse interval is z 0m , as shown below:

[0086]

[0087] Then the pulse interval probability density of all motion pattern areas is calculated, where the probability density Q of the mth rotating disk pattern is um (p 0m ) is as follows:

[0088]

[0089] The similarity function Bhattacharyya coefficient is used to measure the target model and all the turntable patterns R P1 , R P2 , R P3 , …, R Pm , ..., and find the real target R P In addition, since the pulse image sensor has a high frame rate, the position change of the target between two or several frames must be small, that is, the target between two or several frames must remain within a certain distance range L d Therefore, the Bhattacharyya coefficient can be adjusted as follows:

[0090]

[0091] where ρ m is the adjusted Bhattacharyya coefficient value of the mth candidate model. In addition, L d It can be adjusted adaptively and iteratively. First, L d Set a very small initial value, and then adjust the m The value of L is adjusted d If the value of m The maximum value of is 0, then L d The size increases by 1 until ρ m The maximum value of is not 0, as shown below:

[0092]

[0093] The real target R P Then ρ m The largest wheel pattern is:

[0094] R P =R Pm ,if ρ m =max(ρ m ),m∈[1,2,3,…] (16)

[0095] Then update the target model, i.e. R Q =R P .

[0096] (4) After finding the real target area, it is necessary to reconstruct the scene and the target. In order to clearly reconstruct the target and the scene it is in, the original pulse interval is filtered. This paper adopts bilateral filtering as shown below:

[0097]

[0098] Where P′ t (i,j) The pulse interval after filtering, P t (i, j) is the original pulse interval, W p is the normalization coefficient, G σs is the spatial kernel, G σr is the pixel kernel, S is the spatial domain of pixel (i, j), σs and σr are the standard deviations of the spatial kernel and pixel kernel. Then the image is reconstructed as follows:

[0099] G t (i,j)=K / P′ t (i,j),K=255×P min (18)

[0100] Where K is the conversion coefficient from pulse interval to grayscale value, Pmin is the minimum pulse interval in the pulse interval sequence. Figure 5 The whole process of the algorithm is shown.

[0101] The parameter A in formula (4) is generally 20, and the parameter R in formula (5) is D is 10, R B is 2, the neighborhood R in formula (7) ij Generally, it is a circle with a radius of 2 centered at pixel (i, j). The parameter Nbin in equations (11) and (13) is generally 8 to 20. The parameters in equation (17) generally use a 5×5×5 matrix window sequence centered at the filtered pixel for bilateral filtering, and the standard deviation σs and σr are more appropriately 5. Figure 6 The motion detection of a high-speed turntable is shown, from which it can be seen that the method can effectively remove ghosts and holes and improve the integrity of motion detection. Figure 7 The figure shows the tracking result of the pulse image sensor on the high-speed turntable pattern. The blue box in the figure shows the tracking effect of this algorithm, and the red and green boxes show the tracking effects of other algorithms. It can be seen that compared with other algorithms, this algorithm has higher tracking accuracy and tracking stability, and is more suitable for pulse sequence image sensors.

[0102] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A high-speed target tracking method based on a pulse sequence image sensor, characterized in that the steps as follows: (1) Obtaining pulse interval and pulse frequency according to pulse data; (2) Perform motion detection on the moving target and remove ghosts and holes that appear during the motion detection process. The specific steps are as follows: The pulse frequency is used as the pixel value to detect moving targets, and the pixel value of frame A is used to initialize the background model, as shown in formula (4): M A (i,j)={f1(i,j),f2(i,j),…,f A (i,j)} (4) Then calculate the Euclidean distance between the pixel value at the current moment and the background model, and compare the calculated result with the set threshold R D Compared with the threshold R D , then the comparison result r k is 1, which is higher than the threshold R D If the statistical result is greater than the given threshold R, then B , then the detection result of the pixel is B t (i, j) is the background, and the binary grayscale value is 0. Otherwise, it is a moving target, and the binary grayscale value is 255, that is: At the same time, if the pixel is detected as background, the background model needs to be updated, that is, the pixel value is used to randomly replace any sample in the background model of any pixel in the neighborhood of the pixel, as shown in formula (6): where f an (i,j) is the background model M of pixel (i,j) A For any sample in (i,j), f an (i an ,j an ) is any sample in the background model of any pixel in the neighborhood of pixel (i, j); Then remove the ghosts and holes that appear in the motion detection process. For pixel (i, j), the neighborhood R of pixel (i, j) is ij Fixed time period t h The pulse signal within the neighborhood is counted and the pulse density d of the neighborhood is calculated. t (i,j): Among them, H t (x,y) is the neighborhood R ij The pulse data of the inner pixel (x, y), the pulse density range at the actual moving target is R H , then although the binary grayscale value of the ghost is 255, its pulse density must not be within R H Within the range, ghosting is eliminated accordingly, namely: Similarly, since the binary grayscale value of the hole inside the target is 0, its pulse density is the same as that of the actual target, the hole is compensated accordingly, that is: B t (i,j)=255,if B t (i,j)=0and d t (i,j)∈R H (9) (3) Locking and tracking the target: First, lock the target to be tracked from all the moving objects detected by motion, and mark the area where the target is located. Then, use the target area at this time as the target model and calculate the probability density of the pulse interval of the area. After one or several frames, mark the areas where all the moving objects detected by motion are located again, and then calculate the probability density of the pulse interval of each moving object area and compare it with the probability density of the target model. Then, the moving object with the most similar probability density to the target model within a certain distance range is the target to be tracked; finally, use the moving object as the new target model, and use the probability density in the moving object area as the probability density of the target model to update the target model for target tracking in the next frame or several frames; (4) After finding the real target area, bilateral filtering is used to reconstruct the image of the scene and the target.

2. The high-speed target tracking method based on a pulse sequence image sensor as claimed in claim 1, characterized in that: in: (1) Obtain the pulse interval and pulse frequency based on the pulse data. For the pixel (i, j) at time t, first find out the output time Zmax of the most recent pulse data 1 before time t and the output time Zmin of the most recent pulse data 1 after time t, as shown in formula (1): Among them I z (i, j) is the pulse data of pixel (i, j) at time z, and then the original pulse interval P of the pixel at time t t (i, j) is obtained by the difference between these two moments, as shown in formula (2): P t (i,j)=Zmin-Zmax (2) The pulse data is converted into pulse intervals by formula (2), and then the pulse frequency is calculated according to the converted pulse intervals, that is: f t (i,j)=1 / P t (i,j) (3) (4) After finding the real target area, bilateral filtering is used to reconstruct the scene and target image, as shown in formula (17): Where P′ t (i,j) The pulse interval after filtering, P t (i, j) is the original pulse interval, W p is the normalization coefficient, G σs is the spatial kernel, G σr is the pixel kernel, S is the spatial domain of pixel (i, j), σs and σr are the standard deviations of the spatial kernel and the pixel kernel, and then the image is reconstructed as shown in formula (18): G t (i,j)=K / P′ t (i,j),K=255×P min (18) Where K is the conversion coefficient from pulse interval to grayscale value, P min is the minimum pulse interval in the pulse interval sequence.

3. The high-speed target tracking method based on a pulse sequence image sensor as claimed in claim 1, characterized in that: In step (3), the target to be tracked is first locked at the initial time t0, the area where the target is located is marked by the connected domain, and the target area at this time is used as the target model and named R Q , the nth pixel in the area is named p n , whose pulse interval is z n , the center pixel of the area is named p0, and its pulse interval is z0, as shown in formula (10): Then calculate the pulse interval probability density Q of the target model u (p0), as shown below: Where K is the Epannechnikov kernel function, h is the window size of the kernel function, and C is the normalization coefficient, so that b(z n ) is the pulse interval z of the pixels in the area n The histogram quantization value, Nbin is the maximum quantization range, u is the eigenvalue, δ(b(z n )-u) is the Delta function, which is used to determine the quantized value b(z n ) belongs to eigenvalue u, N um is the number of all pixels in the area. After one or several frames, that is, at time t, the area where all moving objects detected by motion are located is marked again through the connected domain and named R P1 , R P2 , R P3 , …, R Pm ,…, assuming that the mth moving object region R Pm The nth pixel in is p nm , whose pulse interval is z nm , the center pixel of the model is p 0m , whose pulse interval is z 0m , as shown in formula (12): Then the pulse interval probability density of all moving object areas is calculated, where the probability density Q of the mth moving object is um (p 0m ) is shown in formula (13): The similarity function Bhattacharyya coefficient is used to measure the target model and all moving objects R P1 , R P2 , R P3 , …, R Pm , ..., and find the real target R P In addition, since the pulse image sensor has a high frame rate, the position change of the target between two or several frames must be small, that is, the target between two or several frames must remain within a certain distance range L d Therefore, the Bhattacharyya coefficient can be adjusted as follows: where ρ m is the adjusted Bhattacharyya coefficient value of the mth candidate model. In addition, L d It can be adjusted adaptively and iteratively. First, L d Set a very small initial value, and then according to ρ m The value of L is adjusted d If the value of m The maximum value of is 0, then L d The size increases by 1 until ρ m The maximum value of is not 0, as shown in formula (15): The real target R P Then ρ m The largest moving object is: R P =R Pm ,if p m =max(ρ m ),m∈[1,2,3,…] (16) Then update the target model, i.e. R Q =R P .

Images