 Research
 Open access
 Published:
Segmentation algorithm via Cellular Neural/Nonlinear Network: implementation on Bioinspired hardware platform
EURASIP Journal on Advances in Signal Processing volume 2011, Article number: 69 (2011)
Abstract
The Bioinspired (Bii) Cellular Vision System is a computing platform consisting of sensing, array sensingprocessing, and digital signal processing. The platform is based on the Cellular Neural/Nonlinear Network (CNN) paradigm. This article presents the implementation of a novel CNNbased segmentation algorithm onto the Bii system. Each part of the algorithm, along with the corresponding implementation on the hardware platform, is carefully described through the article. The experimental results, carried out for Foreman and Carphone video sequences, highlight the feasibility of the approach, which provides a frame rate of about 26 frames/s. Comparisons with existing CNNbased methods show that the conceived approach is more accurate, thus representing a good tradeoff between realtime requirements and accuracy.
1. Introduction
Due to the recent advances in communication technologies, the interest in video contents has increased significantly, and it has become more and more important to automatically analyze and understand video contents using computer vision techniques. In this regard, segmentation is essentially the first step toward many image analysis and computer vision problems [1–15]. With the recent advances in several new multimedia applications, there is the need to develop segmentation algorithms running on efficient hardware platforms [16–18]. To this purpose, in [16] an algorithm for the realtime segmentation of endoscopic images running on a specialpurpose hardware architecture is described. The architecture detects the gastrointestinal lumen regions and generates binary segmented regions. In [17], a segmentation algorithm was proposed, along with the corresponding hardware architecture, mainly based on a connected component analysis of the binary difference image. In [18], a multiplefeatures neuralnetworkbased segmentation algorithm and its hardware implementation have been proposed. The algorithm incorporates static and dynamic features simultaneously in one scheme for segmenting a frame in an image sequence.
Referring to the development of segmentation algorithms running on hardware platforms, in this article the attention is focused on the implementation of algorithms running on the Cellular Neural/Nonlinear Network (CNN) Universal Machine [5–7]. This architecture offers great computational capabilities, which are suitable for complex imageanalysis operations in objectoriented approaches [8–10]. Note that so far few CNN algorithms for obtaining the segmentation of a video sequence into moving objects have been introduced [5, 6]. These segmentation algorithms were only simulated, i.e., the hardware implementation of these algorithms is substantially lacking. Based on these considerations, this article presents the implementation of a novel CNNbased segmentation algorithm onto the Bioinspired (Bii) Cellular Vision System [9]. This system builds on CNN type (ACE16k) and DSP type (TX 6×) microprocessors [9]. The proposed segmentation approach focuses on the algorithmic issues of the Bii platform, rather than on the architectural ones. This algorithmic approach has been conceived with the aim of fully exploiting both the capabilities offered by the Bii system, that is, the analog processing based on the ACE16k as well as the digital processing based on the DSP. We would point out that, referring to the segmentation process, the goal of our approach is to find moving objects in video sequences characterized by almost static background. We do not consider in this article still images or moving objects in a video captured by a camera located on a moving platform, where the background is also moving.
The article is organized as follows. Section 2 briefly revises the basic notions on the CNN model and the Bii cellular vision architecture. Then the segmentation algorithm is described in detail (see the block diagram in Figure 1). In particular, in Section 3, the motion detection is described, whereas Section 4 presents the edge detection phase, which consists of two blocks, the preliminary edge detection and the final edge detection. In Section 5, the object detection block is illustrated. All the algorithms are described from the point of view of their implementation on the Bii, that is, for each task it is specified which templates (of the CNN) run on the ACE16k chip and which parts run on the DSP. Finally, Section 6 reports comparisons between the proposed approach and the segmentation algorithms described in [3] and [5], which have been also implemented on the Bii Cellular Vision System.
2. Cellular Neural/Nonlinear Networks and BioInspired Cellular Vision System
Cellular Neural/Nonlinear Networks represent an information processing system described by nonlinear ordinary differential equations (ODEs). These networks, which are composed of a large number of locally connected analog processing elements (called cells), are described by the following set of ODEs [1]:
where x_{ ij } (t) is the state, y_{ ij } (t) the output, and u_{ ij } (t) the input. The constant I_{ ij } is the cell current, which could also be interpreted as a spacevarying threshold [19]. Moreover, A_{ ij,kl } and B_{ ij,kl } are the parameters forming the feedback template A and the control template B, respectively, whereas kl\in {N}_{\stackrel{\u0304}{r}} is a grid point in the neighborhood within the radius \stackrel{\u0304}{r}of the cell ij[20].
Since the cells cooperate in order to solve a given computational task, CNNs have provided in recent years an ideal framework for programmable analog array computing, where the instructions are represented by the templates. This is in fact the basic idea underlying the CNN Universal Machine [1], where the architecture combines analog array operations with logic operations (therefore named as analogic computing). A global programming unit was included in the architecture, along with the integration of an array of sensors. Moreover, local memories were added to each computing cell [1]. The physical implementations of the CNN Universal Machine with integrated sensor array proved the physical feasibility of the architecture [11, 12].
Recently, a Bioinspired (Bii) Cellular Vision System has been introduced, which combines Analogic Cellular Engine (ACE16k) and DSP type microprocessors [9]. Its algorithmic framework contains several feedback and automatic control mechanisms among the different processing stages [9]. In particular, this article exploits the Bii Version 2 (V2), which has been described in detail in reference [9]. The main hardware building blocks of this Bii architecture are illustrated in Figure 2. It has a color (1280 * 1024) CMOS sensor array (IBIS 5C), two highend digital signal processors (TX C6415 and TX C6701), and a communication processor (ETRAX 100) with some external interfaces (USB, FireWire, and a general digital I/O, in addition to the Ethernet and RS232).
Referring to the Analogic Cellular Engine ACE16k, note that a full description can be found in [12]. Herein, we recall that it represents a low resolution (128 * 128) grayscale image sensor array processor. Thus, the Bii is a reconfigurable device, i.e., it can be used as a monocular or a binocular device with a proper selection of a highresolution CMOS sensor (IBIS 5C) and a lowresolution CNN sensor processor (ACE16k) [9].
Two tools can be used in order to program the Bii Vision System, i.e., the analogic macro code (AMC) and the software development kit (SDK). In particular, by using the AMC language, the Bii Vision System can be programmed for simple analogic routines [9], whereas the SDK is used to design more complex algorithms (see Appendix). Referring to the image processing library (IPL), note that the socalled TACE_IPL is a library developed within the SDK. It contains useful functions for morphological and greyscale processing in the ACE16k chip (see Appendix). Additionally, the Bii V2 includes an InstantVision™ library [9].
Finally, note that through the article, the attention is focused on the way the proposed segmentation algorithm is implemented onto the Bii Cellular Vision System. Namely, each step of the algorithm has been conceived with the aim of fully exploiting the Bii capabilities, i.e., the processing based on the ACE16k chip as well as the processing based on the DSP.
3. Motion detection
This section illustrates the motion detection algorithm (Figure 1). Let {Y}_{i}^{\mathsf{\text{LP}}} and {Y}_{i\mathsf{\text{}}3}^{\mathsf{\text{LP}}} be two graylevel images, processed by a lowpass (LP) filtering, and let {Y}_{i}^{\mathsf{\text{MD}}} be the motion detection (MD) mask. In order to implement the motion detection onto the Bii, the first step (see Equation 3) consists in computing the difference between the current frame {Y}_{i}^{\mathsf{\text{LP}}} and the third preceding frame {Y}_{i\mathsf{\text{}}3}^{\mathsf{\text{LP}}} using the ACE16k chip. The indices i and i3 denote that the frames i2 and i1 are skipped. Namely, the analysis of the video sequences considered through the article suggests that it is not necessary to compute the difference between successive frames, but it is enough every three frames. However, as far as the algorithm goes, every frame is evaluated, even though the reference frame is three frames older. This means that we need to store every frame, because the frame i + 1 requires frame i2 as a reference.
Then, according to Step 2 in Equation 3, positive and negative threshold operations are applied to the difference image via the ConvLAMtoLLM function [13] implemented on the ACE16k chip. This function (included in the SDK) converts a greylevel image stored in the local analog memory (LAM) into a binary image stored in the local logic memory (LLM). Successively, the logic OR operation is applied between the output of the positive threshold and the output of the negative threshold. The resulting image includes all the changed pixels.
Finally, according to Step 3, the Point Remove function[13] (running on the ACE16k) is used for deleting irrelevant pixels not belonging to the contour lines. The output of the algorithm is the MD mask {Y}_{i}^{\mathsf{\text{MD}}}, which entirely preserves the moving objects. Figure 3a, c shows a sample frame of Foreman and Carphone video sequences, respectively, whereas Figure 3b, d shows the corresponding motion detection mask {Y}_{i}^{\mathsf{\text{MD}}}.
4. Edge detection
The proposed edge detection phase consists of two blocks, the preliminary edge detection and the final edge detection (see Figure 1). In the first block, the CNNbased dual window operator (proposed by Grassi and Vecchio [10]) is exploited to reveal edges as zerocrossing points of a difference function, depending on the minimum and maximum values in the two windows. After this preliminary selection of edge candidates, the second block enables accurate edge detection to be obtained, using a technique able to highlight the discontinuity areas.
4.1. Preliminary edge detection
The aim of this phase is to locate the edge candidates. The dual window operator is based on a criterion able to localize the mean point within the transition area between two uniform luminance areas [10]. Thus, the first step consists in determining the minimum and maximum values in the two considered windows. Given the input image {Y}_{i}^{\mathsf{\text{LP}}}, we consider for each sample s\in {Y}_{i}^{\mathsf{\text{LP}}}\left(x,y\right) two concentric circular windows, centered in s and having radius r and R, respectively (r < R). Let M^{R} and m^{R} be the maximum and minimum values of {Y}_{i}^{\mathsf{\text{LP}}} within the window of radius R, and let M^{r} and m^{r} be the maximum and minimum values within the window of radius r[10]. Note that, for the videosequences considered through the article, we have taken the values r = 1 pixel and R = 2 pixels. For each sample s, let us define the difference function D(s) = α_{1} (s)  α_{ 2 } (s), where α_{1} (s) = M^{R}  M^{r} and α_{2} (s) = m^{r}  m^{R} . By assuming that s is the middle point in a luminance transition, the relationship α_{1} (s) = α_{ 2 } (s) holds. In the case of noise, the change in the sign of the difference function D(s) is a more effective indicator of the presence of a contour [10]. Since D(s) approximates the directional derivative of the luminance signal along the gradient direction [10], the relationship D(s) = 0 is equivalent to find the flex points of luminance transitions. In particular, we look for zeropoints and zerocrossing points of D(s). Hence, the introduction of a threshold is required, so that samples s satisfy the condition threshold < D (s) < threshold. Successively, edge samples are detected according to the following algorithm [10]:
In other words, by applying the algorithm (4) to the sample itself and to the four neighboring samples, preliminary edge detection is achieved. In order to effectively implement (4) onto the Bii, the first step is the computation of D(s), which can be realized using orderstatistics filters. They are nonlinear spatial filters that enable maximum and minimum values to be readily computed onto the Bii platform. Their behaviors consist in ordering the pixels contained in a neighborhood of the current pixel, and then replacing the pixel in the centre of the neighborhood with the value determined by the selected method. Therefore, these filters are well suited to find the minimum and maximum values in the neighborhood of the current pixel. The implementation of D(s) gives the images in Figure 4a, c for Foreman and Carphone, respectively.
Going to Step 2, the threshold is implemented on the ACE16k using the ConvLAMtoLLM function. Then, the relationship threshold < D (s) < threshold is satisfied by implementing the operations inversion, OR and inversion again onto the ACE16k chip. Note that we look for samples s so that D(s) = 0. Additionally, we look for samples s satisfying the condition that D(s) ≥ 0 but, simultaneously, D(s) must be negative in a crossshape neighborhood of s. Specifically, at least one of the four conditions D(x_{0} ± 1,y_{0} ± 1) < 0 must be satisfied. Thus, we need to compute D(s) by exploring proper neighborhoods of (x_{0},y_{0}), two examples of which are reported in Figure 4e, f. Note that the object is represented by black pixels, while the background is represented by white pixels. The exploration of proper neighborhoods in the image D(s) can be done using the morphologic dilate4 function, which performs fourconnectivity (crossmask) binary dilatation on the ACE16k [13]. Note that Figure 4e contains an edge, since the conditions D(x_{0}1,y_{0}) < 0 and D(x_{0},y_{0}1) < 0 are satisfied. On the other hand, Figure 4f does not contain any edge, since D(s) > 0 in the neighborhood of (x_{0},y_{0}). Referring to Foreman, the edges selected by implementing the condition threshold < D (s) < threshold are reported in Figure 4g, whereas those selected by exploring proper neighborhoods of (x_{0},y_{0}) are reported in Figure 4h. In particular, note that Figure 4h highlights that there are some flat areas characterized by some edges. Finally, the OR operation between the images in Figure 4g, h provides the image {Y}_{i}^{\mathsf{\text{prel}}} representing the preliminary edge detection. To this purpose, Figure 4b, d depicts the images {Y}_{i}^{\mathsf{\text{prel}}} for Foreman and Carphone video sequences, respectively.
4.2. Final edge detection
The aim of this phase is to better select the previously detected edges. Referring to the previous section, note that the zeros of D(s) are not only flex points of luminance transitions, but also the set of pixels having a neighborhood where luminance is almost constant [10]. Since noise causes small fluctuations, these fluctuations may generate changes in the sign of D that would be incorrectly assumed as edge points. Therefore, in order to better select the edges detected in the previous phase, we need to integrate the available information with the slope of the luminance signal. To this purpose, note that M^{R} and m^{R} identify the direction of maximum slope in the neighborhood of s[10]. Therefore, by suitably exploiting M^{R} and m^{R} , we first need to generate a matrix S, which takes into account the slope of the luminance signal. Then, a threshold gradient operation is applied to S, with the aim to obtain a gradient matrix G. Namely, the final objective is to obtain an image that includes all the edges selected by the gradient operation (i.e., {Y}_{i}^{\mathsf{\text{grad}}}). Successively, the image {Y}_{i}^{\mathsf{\text{grad}}} needs to be cleaned and skeletonized, in order to reduce all the edges to onepixel thin lines. The image reporting the final edge detection, indicated by {Y}_{i}^{\mathsf{\text{finaledge}}}\left(s\right), can be obtained by applying the following algorithm:
In order to effectively implement the algorithm (5) onto the Bii, at first the matrix D(s) is processed by means of the ConvLAMtoLLM function, which implements the threshold 'zero' on D(s). Then, the pixels in D that correspond to D(s) ≥ 0 assume the maximum value of the luminance signal (within the window of radius R) and generate the image {M}_{D}^{R}. Similarly, the pixels in D that correspond to D(s) < 0 assume the minimum value of the luminance signal and generate the image {m}_{D}^{R}. Then, in order to implement the matrix S(s), we need the following new switch template:
The matrix S(s) is generated onto the ACE16k chip, where {M}_{D}^{R} is used as input, {m}_{D}^{R} as state whereas the output of the 'zero' threshold is used as mask. Referring to the template (6), we have chosen the name switch since the image S(s) is obtained by 'switching' between M^{R} (s) and m^{R} (s), depending on the mask values. Note that the template (6), by providing the matrix S(s), enables the slope of the luminance signal to be taken into account. The experimental result of S(s) are reported in Figures 5a and 6a for Foreman and Carphone, respectively.
Then, according to the algorithm (5), we need to implement the threshold gradient operation onto the Bii. This can be done using a sequence of eight templates, applied in eight directions N, NW, NE, W, E, SW, S, and SE. For example, referring to the NW direction, the following novel template is implemented on the ACE16k:
where the bias is used as a threshold level (herein, thres = 1.1). The other seven remaining templates can be easily derived from (7). Then the logic OR is applied to the eight output images in order to obtain a single image, which is denoted by G(s) (see Figure 5b). Note that G stands for gradient, given that it represents the output of the threshold gradient (7). However, the image G needs to be cleaned, since it usually contains some open lines (see the upper leftside in Figure 5b). These open lines can be deleted by applying the prune template:
The output of the prune function is reported in Figure 5c, where it can be seen that the open line in the upper leftside part has been partially deleted. Note that the prune function also enables the back part in Figure 5c to become more compact (i.e., the white dots in the black part have disappeared). Then, the hollow template reported in [13] has to be applied. This template, running on the ACE16k chip, enables the concave locations of objects to be filled. In order to achieve this objective, the hollow template needs to be applied. The output of the hollow is shown in Figure 5d. The white part in Figure 5d indicates that the corresponding part in the image S(s) does not contain information related to edges. Since the hollow is timeconsuming, it is useful to carry out this operation by exploiting the great computational power offered by the CNN chip.
Finally, by using the switch template (6) with input = {Y}_{i}^{\mathsf{\text{prel}}}\left(s\right), state = ∅ (i.e., the white image) and mask = G (s), it is possible to obtain the image {Y}_{i}^{\mathsf{\text{grad}}}\left(s\right), which includes all the edges selected by the gradient operation (see Figures 5e and 6b). In order to skeletonize {Y}_{i}^{\mathsf{\text{grad}}}\left(s\right) and reduce all the edges to onepixel thin lines, the skeletonization function (included in the TACE_IPL library) is implemented on the ACE16k chip. Then, in order to complete open edges (if any) we can use the dilation and erosion functions included in the TACE_IPL. Specifically, we first apply the dilation function, and then the erosion function. These two functions are applied from three to six times, depending on the video sequence under consideration. Finally, the last step lies in deleting the remaining open lines. By applying the prune template (8), the final edges can be obtained, as shown by the images {Y}_{i}^{\mathsf{\text{finaledge}}}\left(s\right) reported in Figures 5f and 6c for Foreman and Carphone, respectively.
5. Object detection
The proposed object detection phase can be described using the following iterative procedure:
First, the following holefiller template is implemented on the ACE16k:
This template is applied to the inverted image of {Y}_{i}^{\mathsf{\text{finaledge}}} with the aim to fill all the holes. Figure 7 depicts the outputs of the holefiller after different processing times, with the aim to show the system behavior when the processing times are increased. Note that the holefiller has to be applied in a recursive way, in order to fill more and more holes. However, differently from Figure 7 that has an explanatory purpose, we need to apply this template by slowly increasing the processing times. Namely, if we slowly increase the processing times, it is possible to highlight at the most two closed objects at a time, so that these objects can be extracted in the next steps. As a consequence, the holefiller plays an important role: by slowly filling the holes in a morphological way, it enables the closed objects to be extracted in the next steps of the algorithm.
In order to implement the second step, the logic XOR is applied between the output of the holefiller (i.e., {Y}_{i}^{\mathsf{\text{fill(}}k\mathsf{\text{)}}}) and the inverted image of {Y}_{i}^{\mathsf{\text{finaledge}}}. Note that the logic XOR enables changes in the two images to be detected. This logic function returns a 1 only if both operands are logically different, otherwise it returns a 0. Bitwise logic XOR is executed on the ACE16k between LLM1 and LLM2 (binary images stored in the Local Logic Memories 1 and 2). Herein, the outcome of the XOR is the binary image {Y}_{i}^{\mathsf{\text{changes(}}k\mathsf{\text{)}}}, which locates the changes between the two images {Y}_{i}^{\mathsf{\text{fill(}}k\mathsf{\text{)}}} and \left({Y}_{i}^{\mathsf{\text{finaledge}}}\right). The output of the XOR is shown in Figure 8a.
According to Step 3, the holefiller template is applied to {Y}_{i}^{\mathsf{\text{fill(}}k\mathsf{\text{)}}}, with the aim to obtain {Y}_{i}^{\mathsf{\text{fill(}}k+1\mathsf{\text{)}}}. Referring to Step 4, the morphologic dilate function is utilized to thicken the contours within the image {Y}_{i}^{\mathsf{\text{fill(}}k+1\mathsf{\text{)}}}. The result of the dilate function, which performs binary dilatation onto the ACE16k, is indicated by {Y}_{i}^{\mathsf{\text{dilation(}}k+1\mathsf{\text{)}}} and is shown in Figure 8b.
According to Step 5, we need to detect the remaining objects in {Y}_{i}^{\mathsf{\text{dilation(}}k+1\mathsf{\text{)}}}. This can be done using the recall template
where the image {Y}_{i}^{\mathsf{\text{dilation(}}k+1\mathsf{\text{)}}} is used as input and the image {Y}_{i}^{\mathsf{\text{finaledge}}} as state. In order to show how the recall template works, Figure 9 shows its output after different processing times. Note that the recall template has to be applied in a recursive way. In particular, by increasing the processing times, note that more and more objects are recalled (see Figure 9).
However, differently from Figure 9 that has an explanatory purpose, herein we need to apply this template by slowly increasing the processing times. Namely, in order to guarantee a satisfying total frame rate, we need to recall few objects at a time, so that the processing times due to the recall template are not large. In this way, the slow recursive application of the recall template does not affect the overall system performances. In conclusion, the recall template plays an important role: by taking into account the image containing the final edge (state), it enables the objects enclosed in the dilated image (input) to be recalled and subsequently extracted.
Now, by applying the recall template (11) using the image in Figure 8b as input and the image in Figure 5f as state, the image reported in Figure 10a is obtained. This image, indicated by {Y}_{i}^{\mathsf{\text{recall(}}k+1\mathsf{\text{)}}}, is constituted by groups of objects. In order to obtain new objects at each iteration, we need to detect the changes between the images {Y}_{i}^{\mathsf{\text{recall(}}k+1\mathsf{\text{)}}} and {Y}_{i}^{\mathsf{\text{changes(}}k\mathsf{\text{)}}}, as indicated by Step 6. To this purpose, we can apply the logic XOR between {Y}_{i}^{\mathsf{\text{recall(}}k+1\mathsf{\text{)}}} and {Y}_{i}^{\mathsf{\text{changes(}}k\mathsf{\text{)}}}. If changes are detected, we need to check whether the extracted object belongs to the moving objects. This operation is implemented by exploiting the AND operation between the output of previous XOR and the motion detection mask {Y}_{i}^{\mathsf{\text{MD}}}. The output of the AND is indicated by {Y}_{i}^{\mathsf{\text{extracted(}}k\mathsf{\text{+1)}}}. For example, the objects extracted after the first iteration are shown in Figure 10b. Finally, the extracted object {Y}_{i}^{\mathsf{\text{extracted(}}k\mathsf{\text{+1)}}} is used to update the image {Y}_{i}^{\mathsf{\text{changes(}}k\mathsf{\text{)}}}, with the aim of obtaining {Y}_{i}^{\mathsf{\text{changes(}}k+1\mathsf{\text{)}}}. This iterative procedure is carried out until all the objects are extracted. Namely, the procedure ends when the condition {Y}_{i}^{\mathsf{\text{fill(}}k\mathsf{\text{)}}} = {Y}_{i}^{\mathsf{\text{fill(}}k+1\mathsf{\text{)}}} is achieved for two consecutive iterations. Figures 8 and 10 summarize some of the fundamental steps of the object detection algorithm for Foreman video sequence. Similar results have been obtained for Carphone video sequence.
6. Discussion
We discuss the results of our approach by making comparisons with previous CNNbased methods illustrated in [3] and [5]. We would remark that the comparison between the proposed approach and the methods in [3] and [5] is homogeneous, since we have implemented all these techniques on the same hardware platform (i.e., the Bii). At first, we compare these approaches by visual inspection. By analyzing the results in Figures 11 and 12, it can be noticed that the proposed technique provides more accurate segmented objects than the ones obtained by the techniques in [5] and [3]. For example, the analysis of Figure 11a suggests that the proposed approach is able to detect man's mouth, eyes, and nose. Note the absence of open lines too. The methods depicted in Figure 11b, c do not offer similar capabilities. Referring to Figure 12a, note that we have obtained an accurate result, since man's mouth, eyes, and nose are detected, along with some moving parts in the back of the car. Again, the approaches depicted in Figure 12b, c do not reach similar performances. It can be concluded that, by exploiting the proposed approach, the edges are much more close to the real edges with respect to the method in [5] and [3].
Now an estimation of the processing time achievable by the proposed approach is given in Table 1. Note that the motion detection and the object detection phases can be fully implemented onto the ACE16k chip, whereas the edge detection phase requires that some parts be implemented on the DSP (see Section 4). The sum of the processing times of the different phases is 37767 μs, which gives a frame rate of about 26 frames/s.
Note that the computational load is mainly due to the DSP in the edge detection phase (28778 μs) and, specifically, to the presence of the orderstatistics filters. On the other hand, these filters are requested to implement the dual window operator, which is in turn required to achieve accurate edge detection, as explained in [10]. Namely, edge detection is a crucial step for segmentation. If we detect edge accurately, we can segment the images correctly. If we analyze the result in reference [5], we note that the authors use a threshold gradient algorithm, which is not particularly suitable for edge detection. On the other hand, the dual window operator is one of the best edge detector (see [10]), even though its implementation is time consuming. Referring to the processing times measured on the Bii for the methods in [3] and [5], their values are 13861 and 5254 μs, respectively. The corresponding frame rates are 72 and 190 frames/s, respectively, while our approach gives 26 frames/s. Thus, the segmentation methods in [3] and [5] are faster than the proposed approach, even though they are less accurate, as confirmed by Figures 11 and 12. Anyway, we believe that 26 frames/s can be considered a satisfying frame rate achievable by the proposed approach, since it represents a good tradeoff between accuracy and speed.
Finally, we would point out that, while we have conducted this research, a novel Bioinspired architecture called EyeRIS vision system has been introduced [21]. It is based on the QEye chip [21], which represents an evolution of the ACE family with the aim to overcome the main drawbacks of ACE chips, such as lack of robustness and large power consumption. Our plan is to implement the segmentation algorithm developed herein on the EyeRIS vision system in the near future. To this purpose, note that one of the authors (F. Karabiber) has already started to work on the EyeRIS vision system, as is proof by the results published in [22].
7. Conclusion
This article has presented the implementation of a novel CNNbased segmentation algorithm onto a Bioinspired hardware platform, called Bii Cellular Vision System [9]. This platform combines the analog processing based on the ACE16k processor [11] as well as the digital processing based on the DSP. The proposed experimental results, carried out for some benchmark video sequences, have shown the feasibility of the approach, which provides a satisfying frame rate of about 26 frames/s. Finally, comparisons with the CNNbased techniques in [5] and [3] have highlighted the accuracy of the proposed method.
Appendix
The software development kit (SDK) is a set of C++ libraries to be used for Bii programming. Some parts of the SDK are based on classes defined in the BaseData module of the InstantVision™ libraries. The SDK is designed to be used together with Code Composer Studio from Texas Instruments (http://www.ti.com/).
The TACE_IPL is an image processing library (IPL) for ACE16k. It contains two function groups for processing images: morphological operations and gray scale operations. The constructor of this class initializes the needed instruction group and writes corresponding IPL templates to the ACE16k.
Note that all the details about the SDK, the InstantVision™ libraries and the TACE_IPL can be found at: http://www.analogiccomputers.com/Support/Documentation/
Alternatively, the Bii programming guide (which includes the SDK and the TACE_IPL) can be requested at: giuseppe.grassi@unisalento.it
Abbreviations
 AMC:

Analogic Macro Code
 Bii:

Bioinspired
 CNN:

Cellular Neural/Nonlinear Network
 IPL:

image processing library
 LP:

low pass
 LAM:

local analog memory
 LLM:

local logic memory
 MD:

motion detection
 ODEs:

ordinary differential equations
 SDK:

software development kit.
References
Chua LO, Roska T: Cellular Neural Networks and Visual ComputingFoundations and Applications. (Cambridge University Press, 2002), Cambridge, UK; ISBN 0521652472
Chen CY, Wang JC, Wang JF, Hu YH: motion entropy feature and its applications to eventbased segmentation of sports video. EURASIP J Adv Signal Process 2008., 8: (Article ID 460913)
Arena P, Basile A, Bucolo M, Fortuna L: An objectoriented segmentation on analog CNN chip. IEEE Trans CASI 2003,50(7):837846. 10.1109/TCSI.2003.813985
Dewan MAA, Hossain MJ, Chae O: An adaptive motion segmentation for automated video surveillance. EURASIP J Adv Signal Process 2008., 13: (Article ID 187413)
Stoffels A, Roska T, Chua LO: Objectoriented image analysis for verylowbitrate videocoding systems using the CNN Universal Machine. Int J Circuit Theory Appl 1997, 25: 235258. 10.1002/(SICI)1097007X(199707/08)25:4<235::AIDCTA961>3.0.CO;2Q
Stoffels A, Roska T, Chua LO: On objectoriented videocoding using the CNN Universal Machine. IEEE Trans CASI 1996,43(11):948952.
Roska T, RodriguezVazquez A: Towards visual microprocessors. Proc IEEE 2002,90(7):12441257. 10.1109/JPROC.2002.801453
Grassi G, Grieco LA: Objectoriented image analysis using the CNN Universal Machine: new analogic CNN algorithms for motion compensation, image synthesis and consistency observation. IEEE Trans CASI 2003,50(4):488499. 10.1109/TCSI.2003.809812
Zarandy A, Rekeczky C: Bii: a standalone ultra high speed cellular vision system. IEEE Circuit Syst Mag 2005,5(2):3645.
Grassi G, Sciascio E Di, Grieco LA, Vecchio P: New objectoriented segmentation algorithm based on the CNN paradigm. IEEE Trans CASII 2006,53(4):259263.
Linan G, Espejo S, DominguezCastro R, RodriguezVazquez A: ACE4k: an analog I/O 64x64 visual microprocessor chip with 7bit analog accuracy. Int J Circuit Theory Appl 2002, (30:):89116.
RodriguezVazquez A, LinanCembrano G, Carranza L, RocaMoreno E, CarmonaGalan R, JimenezGarrido F, DominguezCastro R, Meana SE: ACE16k: the third generation of mixedsignal SIMDCNN ACE chips toward VSoCs. IEEE Trans CASI 2004,51(5):851863. 10.1109/TCSI.2004.827621
Ahn JK, Lee DY, Lee C, Kim CS: Automatic moving object segmentation from video sequences using alternate flashing system. EURASIP J Adv Signal Process 2010., 14: (Article ID 340717)
Hsu CY, Yang CH, Wang HC: Multithreshold level set model for image segmentation. EURASIP J Adv Signal Process 2010., 8: (Article ID 950438)
Kim J, Chen T: A VLSI architecture for videoobject segmentation. IEEE Trans CAS Video Technol 2003,13(1):8396. 10.1109/TCSVT.2002.808082
Ranganathan N, Mehrotra R: A VLSI architecture for dynamic scene analysis. Comp Vis Graph Image Process 1991, 5: 189197.
Kim J, Chen T: Multiple feature clustering for image sequence segmentation. Pattern Recog Lett 2001, 22: 12071217. 10.1016/S01678655(01)000538
Brucoli M, Carnimeo L, Grassi G: A global approach to the design of discretetime cellular neural networks for associative memories. Int J Circuit Theory Appl 1996,24(4):489510. 10.1002/(SICI)1097007X(199607/08)24:4<489::AIDCTA930>3.0.CO;2F
Grassi G: On discretetime cellular neural networks for associative memories. IEEE Trans CASI 2001,48(1):107111. 10.1109/81.903193
RodríguezVázquez A, DomínguezCastro R, JiménezGarrido F, Morillas S, Listán J, Alba L, LiñánCembrano G, Carranza L: The EyeRIS CMOS vision system. In Analog Circuit Design. Berlin, Germany; 2007:1532.
Karabiber F, Arena P, Fortuna L, Fiore S De, Vagliasindi S, Arik S: Implementation of a moving target tracking algorithm using EyeRIS Vision System on a mobile robot. J Signal Process Syst 2010.
Author information
Authors and Affiliations
Corresponding author
Additional information
8. Competing interests
The authors declare that they have no competing interests.
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 2.0 International License (https://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
About this article
Cite this article
Karabiber, F., Vecchio, P. & Grassi, G. Segmentation algorithm via Cellular Neural/Nonlinear Network: implementation on Bioinspired hardware platform. EURASIP J. Adv. Signal Process. 2011, 69 (2011). https://doi.org/10.1186/16876180201169
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/16876180201169