An innovative chain coding mechanism for information processing and compression using a virtual bat-bug agent-based modeling simulation

Highlights

• Developing a hypothetical bat-bug agent-based model to be used as a chain code.

• The obtained compression ratios outperformed standardized benchmarks including JBIG2.

• The design employs a relative encoding technique for enhanced compression.

Abstract

The continuous changes in the size of data create new challenges to design new techniques to reduce its size and encode it in a way that changes its original representation. In this article, we develop a bat-bug agent-based modeling simulation for chain coding and employ it in compressing bi-level image information. The system consists of agents that are classified into static and dynamic depending on their movements. Bugs are considered static agents, and they are distributed over the virtual environment according to the allocation of pixels in the original image. On the other hand, bats are dynamic agent, and their role is to move around to consume bugs while the algorithm tracks their movements. Bats are designed in a way to move within certain boundaries to avoid crashing into each other. Bats employ specific movements that allow them to move in relative directions. Therefore, the frequency of their movements can follow a certain pattern that can help in further size reduction. In other words, the integration of relative movements into our design proved to be advantageous because there is an observable pattern of repeated movements, which allows getting higher compression results. Finally, arithmetic coding is applied to the final strings that represent the movements of bats while searching for bugs to eat. To assess the performance of the algorithm, we compared the findings against standardized benchmarks used in the image processing community: G3, G4, JBIG1, and JBIG2. The outcomes show that we could outperform all these benchmarks using all the images we used for testing. Additionally, we conducted a series of paired samples t-tests, and they revealed that the mean differences between our results and those obtained from other benchmarks are statistically significant.

Introduction

Recent years have witnessed a growing interest in the field of Artificial Intelligence (AI) and its applications in numerous disciplines. The most significant developments are those related to handling big data applications that continue to arise due to the appearance of many platforms such as the Internet of Things (IoT) and Smart Cities (SC) (Han and Wang, 2021, Nguyen et al., 2021). These platforms and many others significantly contributed to the accelerated growth in the volumes of digital information. This tremendous increase in digital information makes transferring large amounts of data over media channels a challenging task. Therefore, it is crucial to investigate new techniques to reduce the volumes of information exchanged between different channels. Agent-based modeling is one of the most inspiring advancements in AI, in general, and data reduction and transformation, in particular. Agents are considered fundamental aspects of agent-based modeling and simulation, and they exhibit specific behaviors during the execution of the program. The behavior of agents within the system allows understanding the system as a whole.

A great deal of previous agent-based modeling research has revealed numerous employments of biological abstractions. It is essential to mention that agent-based modeling is not the only area of research that utilizes these abstractions. Extensive research shows that a significant amount of studies that involve biological abstractions are designed to handle optimization problems (e.g. Okur and Altan (2021), Altan (2020), Sezer and Altan (2021) and Altan and Parlak (2020)). Fig. 1 shows the classification of papers that utilize biological abstractions in agent-based modeling and optimization research. Although the behavior of agents has been applied to various problems in agent-based modeling and optimization, our considerable research reveals that our group is taking the lead in developing a new direction of studies that employ biological abstractions in encoding bi-level image information (e.g. Dhou and Cruzen, 2021a, Dhou and Cruzen, 2019 and Dhou (2019a)). This new direction considers an interdisciplinary approach that combines bio-inspired computational techniques and chain coding principles to develop new agent-based models that encode bi-level image information. Classical image encoding began with Freeman Chain Code which considers moving around the contours and encoding each step (Freeman, 1961). This approach of encoding utilizes one of 4 or 8 codes to represent each direction of contour movements. Our new line of research involves the utilization of biological abstractions embedded with agent-based modeling simulations to develop new chain codes for compression. For example, the ant colony approach (Mouring et al., 2018) applies some abstractions that exist in biological ants, including the pheromone to track the information during the encoding process. On the other hand, the wolf–sheep predation method (Dhou, 2018a, Dhou, 2020a) includes simulations that consist of wolves and sheep where wolves have specific movements used in traversing certain parts of the virtual worlds representing actual images. Furthermore, Dhou and Cruzen (Dhou and Cruzen, 2019) developed an agent-based modeling simulation that takes advantage of the natural reproduction abstraction by allowing existing agents to reproduce additional agents for further information processing. All these agent-based modeling approaches proved to be effective in applying biological abstractions in image coding and compression.

The motivation of this research is to extend the previous agent-based modeling studies applied in encoding bi-level image information. While existing agent-based models proved to be successful in capturing image information, some of them have limitations related to the distribution of the agents and the way they encode information. For instance, the main limitation of the biological reproduction approach (Dhou and Cruzen, 2019) is the ability of some agents to block other agents. Similarly, the ant colony method (Mouring et al., 2018) can lead to having more ants in some image areas as they follow the pheromone to identify pieces of information. Depending on the encountered scenario, this feature can slow the process and result in poor agent distribution over a virtual environment used as image representation. All these limitations were addressed in the present model by creating a bat-bug agent-based modeling simulation and employing it in developing a new chain code to be used in bi-level data compression.

In the current model, an image is first transformed into a virtual environment consisting of bugs spread over different locations. The allocation of the bugs is based on the distribution of the pixels in the original image, and these locations are considered static. In other words, bugs do not move around during the execution of the program. After that, bats are added to the virtual environment, and they are considered dynamic agents that move around while searching for bugs to eat. Bats have specific movements, and their goal is to eat all the bugs in the virtual environment, which declares the end of the simulation. Each bat searches for bugs within a certain radius, and this eventually allows each bat to work on a particular image segment. The model simulates a navigation system for bats, which allows them to detect other bats and bugs in their neighborhood and thus, avoid crashing into each other. Additionally, the design employs fewer codes, which gives more freedom to include additional behaviors while having the same number of bits to represent each code. After recording the number of moves encountered by dynamic agents, the algorithm applies additional processing to further reduce the length of the chains and make them suitable for different operations, including compression.

All agent-based modeling simulations our research group developed for image processing research handle bi-level images. Although bi-level information looks simple, researchers continuously utilize it in different aspects such as psychological, medical, and computing research projects (e.g. Dhou et al. (2018), Dhou (2013), Dhou et al. (2013), Ksasy et al. (2018) and Padlia and Sharma (2019)). The strength of the model comes from its ability to capture the digital information and convert it to a new form that not only reduces the original size of the data, but also allows it to be represented in another form. We believe that such capacity can make our algorithm attractive to researchers from other disciplines. What makes the proposed method suitable for encoding is that it allows bats to work within certain borders and takes advantage of existing NetLogo features and agents to develop the simulation. In other words, bats are programmed in a similar way as some NetLogo agents, and they can simultaneously work on encoding information. One advantage of the NetLogo agents is that they are practical and allow simulating numerous abstractions to be investigated in the image processing domains, such as the ones in Dhou and Cruzen, 2019, Dhou and Cruzen, 2020. Another advantage is utilizing some features that exist in some chain coding applications, such as the relative encoding, which allows bats to encode based on how a movement is related to another in the neighborhood. Taken together, we believe that the relative encoding approach and employment of agents are all practical features that distinguish our work from many other traditional image processing approaches that do not employ biological abstractions, such as the ones in Liu and Žalik (2005) and Zahir and Dhou (2007). It is important to note that numerous bio-inspired models run in virtual environments, which proved their usefulness in solving many problems, including personality analysis (Dhou, 2018b, Dhou, 2019b, Caci and Dhou, 2020, Dhou, 2020b, Dhou, 2021b, Dhou, 2021a).

The reason why we choose the inspiration from the distribution of bats and bugs mainly relies on how the model is structured and how the information can be represented. In other words, the current structure of the bat-bug agent-based model allows us to create a new image representation that consists of bugs to be traced by bats. While bugs in the present model can accurately represent image information, bats are the moving agents that are used during the encoding process. Additionally, the model is designed in a way that the inclusion of bats allows them to work separately within confined spaces. Therefore, offering an additional advantage over using a traditional encoding mechanism. Most importantly, the current model promotes the usage of a bio-inspired computational method in encoding. A bio-inspired method can be incorporated within an agent-based modeling simulation that makes it flexible in exploring numerous parameters and investigating how they work for images of different characteristics. Additionally, the model makes it easy to integrate image processing algorithms, the purpose of which is to enhance compression. Interestingly, although the utilization of agent-based modeling in solving image processing problems is a relatively new research direction, existing research shows that it is very promising. To put it in another way, researchers built agent-based modeling simulations for numerous purposes (e.g. Asgary et al. (2021) and Valtchev et al. (2021)), however, very few simulations were built to address image processing problems. Existing models used in image processing were effective in capturing numerous behaviors and offering various representations that can be examined for different image sets.

Although the majority of the agent-based modeling studies developed by our group employ bio-inspired computational approaches; they have deep roots in image processing research. In other words, the closest image processing research related to our work is called ‘chain code,’ which encodes the connected components in bi-level images (Freeman, 1961). The main advantages of agent-based modeling approaches over traditional chain coding techniques are the following:

• Agent-based modeling approaches provide much flexibility in terms of integrating bio-inspired abstractions within the chain coding application. Examples of abstractions are pheromone in ants (Mouring et al., 2018), predatory behavior (Dhou, 2020a), echolocation in dolphins (Dhou and Cruzen, 2020), and territories in beavers (Dhou and Cruzen, 2021a).

• Agents have the ability to simultaneously work on multiple parts of an image depending on the abstractions utilized by the programmer. For example, in the ant colony algorithm (Mouring et al., 2018), ants are programmed to employ the pheromone feature that attracts them to certain parts of an image. On the other hand, the biological reproduction method (Dhou and Cruzen, 2019) enables agents to generate more agents that can further work on encoding digital information.

• Some of the existing agent-based modeling approaches that utilize biological abstractions (e.g. Dhou and Cruzen, 2019, Dhou and Cruzen, 2020 and Mouring et al. (2018)) have fewer codes than some classical image processing approaches. For example, the chain coding approaches in Liu and Žalik (2005) and Zahir and Dhou (2007) utilize 8 codes, while some agent-based modeling approaches use less number of codes (e.g. Mouring et al. (2018) and Dhou and Cruzen (2020)). The advantage of using less number of codes lies in the fact that a researcher can add further codes that represent combinations and, therefore reduce the size of the final chains.

It is important to mention that in general, agent-based modeling, which is the methodology adopted in this research study, has no theoretical derivations (Van Dyke Parunak et al., 1998, Marshall, 2017, Axelrod, 2006). In other words, agent-based modeling is a different research direction than equation-based modeling (Van Dyke Parunak et al., 1998, Antunes and Takadama, 2007). However, research shows that it has benefits over other numerical modeling strategies (Marshall, 2017, Garnett et al., 2011): first, agents can have different features that allow simulating complex behaviors; and second, agents communicate with one another to create productive arrangements. Modeling these “connections” allows different types of network analysis that are almost impossible using mathematical techniques. Additionally, agent-based modeling proved to be an applicable approach in many fields such as policy-making (Hadzikadic et al., 2015), stock market (Hessary and Hadzikadic, 2017), biology (Carmichael and Hadzikadic, 2013), and epidemiology  (Dréau et al., 2009). The main contributions of this article are the following:

• The current hypothetical model explores the outcomes of multiple bats working on a virtual world representing a bi-level image. This is advantageous over classical image processing approaches since it allows working on different parts of an image simultaneously. Additionally, the algorithm manages the movements of bats and allows each to work on a particular area. Interestingly, although our implementation has fewer details than JBIG algorithms, it could significantly outperform them, as shown in Section 4.

• This is the first attempt to use a bat-bug model in bi-level image coding and compression. Our model can also be explored for further utilization in image processing applications such as security, indexing, and retrieval (e.g. Tariq et al. (2020), Ehsan and Khan (2012), Khan et al. (2014), Ali and Khan (2013) and Bel and Sam (2021)). In other words, the current design results in generating new chains of data from the movements of dynamic agents. These chains can be further processed and reduced in size to act as new signatures to be utilized by researchers from various domains. Additionally, our algorithm has the flexibility to be applied to transformed images, and this, in turn, can generate signatures that are less in size.

• The movements of each bat are represented by five possible directions. This means there is room to include additional directions of different combinations and would still need three bits to represent each direction. Further analysis of the results can guide on adding more directions based on different code combinations. Alternatively, the current design allows the integration of further behaviors while maintaining the number of codes. Existing research reveals the flexibility of agent-based models to integrate new behaviors for encoding, such as the pheromone in ants, natural reproduction, and many others (Mouring et al., 2018, Dhou and Cruzen, 2019). All these behaviors were added without increasing the original number of relative codes. In other words, we can take advantage of particular behaviors and explore their performance on different image types while maintaining the same number of codes utilized in representation.

• Although bio-inspired computational techniques have been intensively addressed by AI researchers, the extensive search reveals that our research group is leading the research direction in employing bio-inspired computational techniques and agent-based modeling in encoding bi-level image information.

The current paper is structured as follows: In Section 2, we provide an overview of related research that exists in the areas of agent-based modeling and image encoding and analysis; In Section 3, we describe the current design, including the model and environment, and outline the main parts of the bat-bug algorithm; In Section 4, we show our results and use different metrics to compare them with standardized benchmarks in compression; Section 5 concludes the article.