Automatic image cropping algorithms aim to recompose images like human-being photographers by generating the cropping boxes with improved composition quality. Cropping box regression approaches learn the beauty of composition from annotated cropping boxes. However, the bias of annotations leads to quasi-trivial recomposing results, which has an obvious tendency to the average location of training samples. The crux of this predicament is that the task is naively treated as a box regression problem, where rare samples might be dominated by normal samples, and the composition patterns of rare samples are not well exploited. Observing that similar composition patterns tend to be shared by the cropping boundaries annotated nearly, we argue to find the beauty of composition from the rare samples by clustering the samples with similar cropping boundary annotations, i.e., similar composition patterns. We propose a novel Contrastive Composition Clustering (C2C) to regularize the composition features by contrasting dynamically established similar and dissimilar pairs. In this way, common composition patterns of multiple images can be better summarized, which especially benefits the rare samples and endows our model with better generalizability to render nontrivial results. Extensive experimental results show the superiority of our model compared with prior arts. We also illustrate the philosophy of our design with an interesting analytical visualization.

Image translation from human faces to anime ones brings a low-end, efficient way to create animation characters for animation industry. However, due to the significant inter-domain difference between anime images and human photos, existing image-to-image translation approaches cannot address this task well. To solve this dilemma, we propose HyProGAN, an exemplar-guided image-to-image translation model without paired data. The key contribution of HyPro-GAN is that it introduces a novel hybrid and progressive training strategy that expands the unidirectional translation between two domains into the bidirectional intra-domain and inter-domain translation. To enhance the consistency between input and output, we further propose a local masking loss to align the facial features between the human face and the generated anime face. Extensive experiments demonstrate the superiority of HyProGAN against state-of-the-art models.

Image cropping aims to enhance the aesthetic quality of a given image by searching for the good cropping views. One common routine is to score and rank the candidate views by the neural network. The network is expected to discriminate the subtle view-wise differences. However, the image-wise differences and the ambiguity in the annotations render difficulties in discriminating the view-wise differences. To focus on the view-wise differences, we propose a feature spliter to build image-wise and view-wise feature and evaluate the candidate views only based on the view-wise feature. Then, we propose the ranking gain loss that alleviates the ambiguity in annotations to amplify the view-wise differences. The remarkable improvement compared with prior arts on public benchmarks illustrates that the view-wise differences matter in cropping view recommendation.

Generative adversarial networks,(GANs) have been trained to be professional artists able to create stunning artworks such as face generation and image style transfer. In this paper, we focus on a realistic business scenario: automated generation of customizable icons given desired mobile applications and theme styles. We first introduce a theme-application icon dataset, termed AppIcon, where each icon has two orthogonal theme and app labels. By investigating a strong baseline StyleGAN2, we observe mode collapse caused by the entanglement of the orthogonal labels. To solve this challenge, we propose IconGAN composed of a conditional generator and dual discriminators with orthogonal augmentations, and a contrastive feature disentanglement strategy is further designed to regularize the feature space of the two discriminators. Compared with other approaches, IconGAN indicates a superior advantage on the AppIcon benchmark. Further analysis also justifies the effectiveness of disentangling app and theme representations. Our project will be released at: https://github.com/architect-road/IconGAN.

Neural Radiance Field (NeRF) and its variants have exhibited great success on representing 3D scenes and synthesizing photo-realistic novel views. However, they are generally based on the pinhole camera model and assume all-in-focus inputs. This limits their applicability as images captured from the real world often have finite depth-of-field (DoF). To mitigate this issue, we introduce DoF-NeRF, a novel neural rendering approach that can deal with shallow DoF inputs and can simulate DoF effect. In particular, it extends NeRF to simulate the aperture of lens following the principles of geometric optics. Such a physical guarantee allows DoF-NeRF to operate views with different focus configurations. Benefiting from explicit aperture modeling, DoF-NeRF also enables direct manipulation of DoF effect by adjusting virtual aperture and focus parameters. It is plug-and-play and can be inserted into NeRF-based frameworks. Experiments on synthetic and real-world datasets show that, DoF-NeRF not only performs comparably with NeRF in the all-in-focus setting, but also can synthesize all-in-focus novel views conditioned on shallow DoF inputs. An interesting application of DoF-NeRF to DoF rendering is also demonstrated. The source code will be made available at: https://github.com/zijinwuzijin/DoF-NeRF.

We show that explicit modeling of composition rules benefits image cropping. Image cropping is considered a promising way to automate aesthetic composition in professional photography. Existing efforts, however, only model such professional knowledge implicitly, e.g., by ranking from comparative candidates. Inspired by the observation that natural composition traits always follow a specific rule, we propose to learn such rules in a discriminative manner, and more importantly, to incorporate learned composition clues explicitly in the model. To this end, we introduce the concept of the key composition map (KCM) to encode the composition rules. The KCM can reveal the common laws hidden behind different composition rules and can inform the cropping model of what is important in composition. With the KCM, we present a novel cropping-by-composition paradigm and instantiate a network to implement composition-aware image cropping. Extensive experiments on two benchmarks justify that our approach enables effective, interpretable, and fast image cropping.

We show that relation modeling between visual elements matters in cropping view recommendation. Cropping view recommendation addresses the problem of image recomposition conditioned on the composition quality and the ranking of views (cropped sub-regions). This task is challenging because the visual difference is subtle when a visual element is reserved or removed. Existing methods represent visual elements by extracting region-based convolutional features inside and outside the cropping view boundaries, without probing a fundamental question: why some visual elements are of interest or of discard? In this work, we observe that the relation between different visual elements significantly affects their relative positions to the desired cropping view, and such relation can be characterized by the attraction inside/outside the cropping view boundaries and the repulsion across the boundaries. By instantiating a transformer-based solution that represents visual elements as visual words and that models the dependencies between visual words, we report not only state-of-the-art performance on public benchmarks, but also interesting visualizations that depict the attraction and repulsion between visual elements, which may shed light on what makes for effective cropping view recommendation.

Image cropping is a common way to enhance the aesthetic quality of images. Huge industrial demand and the tediousness of image cropping make automatic image cropping a prosperous task. Existing works, however, face two difficulties: objects are easily truncated and key components of images are discarded by the model. The key to solving this problem is to understand the relations between different components of an image. These relations break the limit of spatial distance and reflect the contextual information in images, which help the model decide whether to retain a component. Motivated by this, a patch-related graph module is proposed to model the relations between different patches of an image. The patch-related features are extracted by a graph convolution layer and then fused with the original local features by a proposed gated unit. Moreover, a gradient layer is designed to embed the edge information in the input. The edge-prior input helps the model read the contents of images and reserve the main objects completely. Experimental results show that our model grasps the inter-patch relations well and performs competitively with other state-of-the-art approaches.

Image cropping aims to enhance the aesthetic quality of a given image by removing unwanted areas. Existing image cropping methods can be divided into two groups: candidate-based and candidate-free methods. For candidate-based methods, dense predefined candidate boxes can indeed cover good boxes, but most candidates with low aesthetic quality may disturb the following judgment and lead to an undesirable result. For candidate-free methods, the cropping box is directly acquired according to certain prior knowledge. However, the effect of only one box is not stable enough due to the subjectivity of image cropping. In order to combine the advantages of the above methods and overcome these shortcomings, we need fewer but more representative candidate boxes. To this end, we propose FCRNet, a fully convolutional regression network, which predicts several context-aware cropping boxes in an ensemble manner as candidates. A multi-task loss is employed to supervise the generation of candidates. Unlike previous candidate-based works, FCRNet outputs a small number of context-aware candidates without any predefined box and the final result is selected from these candidates by an aesthetic evaluation network or even manual selection. Extensive experiments show the superiority of our context-aware candidates based method over the state-of-the-art approaches.

Understanding the emergence of cooperation among selfish individuals is an enduring conundrum in evolutionary biology, which has been studied using a variety of game theoretical models. Most of the previous studies presumed that interactions between individuals are discrete, but behavior in real systems can hardly be expected to have this dramatically discrete nature. In addition, existing research on continuous strategy games mostly focus on infinite well-mixed populations. Especially, there is few theoretical work on their evolutionary dynamics in structured populations. In the previous work [1], we theoretically studied the game dynamics of continuous strategies in a spatially structured population with its average degree k ≥ 3 under weak selection. Here, we study their evolutionary dynamics under weak selection on a cycle (k = 2), where each individual only interacts with its two immediate neighbors. Using the concept of fixation probability, we derive exact conditions for natural selection favoring one strategy over another for three update rules, called `birth-death', `death-birth', and `imitation'. It shows that for continuous strategy games, the same conditions are derived; especially, the simple rule b/c >; k is valid as well, where b/c is the benefit-to-cost ratio of an altruistic act. In addition, we present a network gain decomposition of the game equilibrium, which might provide a new view of network reciprocity, one of five mechanisms for evolution of cooperation.

One of the major challenges in the field of evolutionary algorithms (EAs) is to characterise which kinds of problems are easy and which are not. Researchers have been attracted to predict the behaviour of EAs in different domains. We introduce fitness landscape networks (FLNs) that are formed using operators satisfying specific conditions and define a new predictive measure that we call motif difficulty (MD) for comparison-based EAs. Because it is impractical to exhaustively search the whole network, we propose a sampling technique for calculating an approximate MD measure. Extensive experiments on binary search spaces are conducted to show both the advantages and limitations of MD. Multidimensional knapsack problems (MKPs) are also used to validate the performance of approximate MD on FLNs with different topologies. The effect of two representations, namely binary and permutation, on the difficulty of MKPs is analysed.

Cooperation is a fundamental principle of all biological systems. Most previous studies presumed that the interactions between individuals are discrete, namely, each individual offers either cooperation or defection. This discrete strategy seems unrealistic in real systems and cooperative behavior in nature should be viewed as a continuous trait. Existing research work on games with a continuous strategy mainly focuses on infinite well-mixed populations. Additionally, our previous work showed that there is a considerable difference in terms of equilibria between continuous and discrete strategy games on graphs under strong selection. This paper studies the game dynamics in finite structured populations under weak selection using the stochastic dynamics based on respectively the mutant fixation probability (ργ) and the fixation probability ratio of mutant to resident (ργ/ρχ). For three update rules, called 'birth death' (BD), 'death-birth' (DB) and 'imitation' (IM), we derive exact conditions for natural selection favoring one strategy over another. Comparing discrete strategy games, we find that for continuous ones (i) the rule, b/c >; k, is also valid; (ii) the same selection conditions are also derived using ργ/ρχ; however, (iii) the selection conditions obtained using ργ and ργ/ρχ are the same instead of different; and (iv) interestingly, the '1/3' rule is not observed for DB and IM updating.

Understanding complex networks in the real world is a nontrivial task. In the study of community structures we normally encounter several examples of these networks, which makes any statistical inferencing a challenging endeavor. Researchers resort to computer-generated networks that resemble networks encountered in the real world as a means to generate many networks with different sizes, while maintaining the real-world characteristics of interest. The generation of networks that resemble the real world turns out in itself to be a complex search problem. We present a new rewiring algorithm for the generation of networks with unique characteristics that combine the scale-free effects and community structures encountered in the real world. The algorithm is inspired by social interactions in the real world, whereby people tend to connect locally while occasionally they connect globally. This local-global coupling turns out to be a powerful characteristics that is required for our proposed rewiring algorithm to generate networks with community structures, power law distributions both in degree and in community size, positive assortative mixing by degree, and the rich-club phenomenon.

Both separated and overlapping communities are useful to analyze real networks in different situations. However, to the best of our knowledge, existing community detection methods based on Evolutionary Algorithms (EAs) can detect separate communities only. This is because it is difficult to represent overlapping communities in ways that are suitable for EAs. In this paper, we first design a representation method that can represent each individual as both separated and overlapping communities without assigning the number of communities in advance. We then design three objective functions to guide the evolutionary process in different conditions. Finally, based on the designed representation and objective functions, we propose a multiobjective evolutionary algorithm to solve CDPs (MEA_CDPs) under the framework of NSGA-II. In the experiments, 4 well-known real-life benchmark networks are used to validate the performance of MEA_CDPs, and the results shown that MEA_CDPs not only can find high quality communities, but also can detect both separated and overlapping communities at the same time, and present multiple types of communities. Moreover, the overlapping nodes identified by MEA_CDPs are really ambiguous according to their edge distributes in different communities. This illustrates the effectiveness of the objective functions we designed.

Based on our previous works, multiagent systems and evolutionary algorithms (EAs) are integrated to form a new algorithm for combinatorial optimization problems (CmOPs), namely, MultiAgent EA for CmOPs (MAEA-CmOPs). In MAEA-CmOPs, all agents live in a latticelike environment, with each agent fixed on a lattice point. To increase energies, all agents compete with their neighbors, and they can also increase their own energies by making use of domain knowledge. Theoretical analyses show that MAEA-CmOPs converge to global optimum solutions. Since deceptive problems are the most difficult CmOPs for EAs, in the experiments, various deceptive problems with strong linkage, weak linkage, and overlapping linkage, and more difficult ones, namely, hierarchical problems with treelike structures, are used to validate the performance of MAEA-CmOPs. The results show that MAEA-CmOP outperforms the other algorithms and has a fast convergence rate. MAEA-CmOP is also used to solve large-scale deceptive and hierarchical problems with thousands of dimensions, and the experimental results show that MAEA-CmOP obtains a good performance and has a low computational cost, which the time complexity increases in a polynomial basis with the problem size.

A novel evolutionary algorithm, Organizational Evolutionary Algorithm for SATisfiability problems (OEA_SAT), is proposed in this paper. OEA_SAT first divides a SAT problem into several sub-problems, and each organization is composed of a sub-problem. Thus, three new evolutionary operators, namely the self-learning operator, the annexing operator and the splitting operator are designed with the intrinsic properties of SAT problems in mind. Furthermore, all organizations are divided into two populations according to their fitness. One is called best-population, and the other is called non-best-population. The idea behind OEA_SAT is to solve the sub-problem first, and then synthesize the solution for the original problem by adjusting the variables which have conflicts. Since the dimensions of sub-problems are smaller and the sub-ones are easy to be solved compared with the original one, the computational cost is reduced in this way. In the experiments, 3700 benchmark SAT problems in SATLIB are used to test the performance of OEA_SAT. The number of variables of these problems is ranged from 20 to 250. Moreover, the performance of OEA_SAT is compared with those of two well-known algorithms, namely WalkSAT and RFEA2. All experimental results show that OEA_SAT has a higher success ratio and a lower computational cost. OEA_SAT can solve the problems with 250 variables and 1065 clauses by only 1.524 seconds and outperforms all the other algorithms.

This paper analyzes the numerical optimization problems from the viewpoint of multiagent systems. First, Macro-Agent Evolutionary Model (MacroAEM) is proposed with the intrinsic properties of decomposable functions in mind. In this model, a subfunction forms a macro-agent, and 3 new behaviors, namely competition, cooperation, and selfishness, are developed for macro-agents to optimizing objective functions. Second, MacroAEM model is integrated with multiagent genetic algorithm, which results a new algorithm, Hierarchical MultiAgent Genetic Algorithm (HMAGA). The convergence of HMAGA is analyzed theoretically and the results show that HMAGA converges to the global optima. In experiments, HMAGA is applied to a kind of complicated decomposable function, namely Rosenbrock function. The results show that HMAGA achieves a good performance, especially for the high-dimensional functions. In addition, the analyses on time complexity demonstrate that HMAGA has a good scalability.

In many applications, one class of data is presented by a large number of examples while the other only by a few. For instance, in our previous works on identification of peer-to-peer (P2P) Internet traffics, we observed that only about 30% of examples can be labeled as ldquoP2Prdquo using a port-based heuristic rule, and even fewer examples can be labeled in the future as more and more P2P applications use dynamic ports. In this paper, the effect of three resampling techniques on balancing the class distribution in training C4.5 and neural networks for identifying P2P traffic is studied. The experimental data were captured at our campus gateway. Nine datasets with different percentages of ldquoP2Prdquo examples and six datasets of different sizes with an actual percentage of about 30% of ldquoP2Prdquo examples are used in the experiments. The results show that resampling techniques are effective and stable, and random over-sampling is a quite good choice for P2P traffic identification considering a combination of the classification performance and time complexity.

We apply streaming data mining techniques, and in particular, concept-adapting very fast decision tree (CVFDT) to identify peer-to-peer (P2P) applications in Internet traffic, as the Internet data flows dynamically in large volumes (streaming data), and in P2P applications, new communities of peers often attend and old communities of peers often leave, requiring the identification methods to be capable of coping with concept drift, and updating the model incrementally. We captured Internet traffic at a main gateway router, performed pre-processing on the captured data, selected the most significant attributes, and prepared a training data stream to which the CVFDT model was applied. We tested our approach on a data stream with 3.5 million P2P and NonP2P traffic records. The results show that our approach can effectively deal with dynamic nature of streaming data and detect the changes in communities of peers. The classification accuracy is higher than 95%, and the method is well-scalable in both time and space complexities, making it competent for large-scale dynamic data. We extracted attributes only from the IP layer, eliminating the privacy concern associated with the techniques that use deep packet inspection.

A new nonslicing floorplan representation, the moving block sequence (MBS), is proposed in this paper. Our idea of the MBS originates from the observation that placing blocks on a chip has some similarities to playing the game, Tetrisreg. Since no extra constraints are exerted on solution spaces, the MBS is not only useful for evolutionary algorithms, but also for dealing with rectangular, convex rectilinear, and concave rectilinear blocks, similarly and simultaneously, without partitioning rectilinear blocks into subblocks. This is owed to a special structure designed for recording the information of both convex and concave rectilinear blocks in a uniform form. Theoretical analyses show that the computational cost of transforming an MBS to a floorplan with rectangular blocks, in terms of the number of blocks, is between linear and quadratic. Furthermore, as a follow-up of our previous works, a new organizational evolutionary algorithm (OEA) based on the MBS (MBS-OEA) is proposed. With the intrinsic properties of the MBS in mind, three new evolutionary operators are designed in the MBS-OEA. To test the performance of the MBS-OEA, benchmarks with hard rectangular, soft rectangular, and hard rectilinear blocks are used. The number of blocks in these benchmarks varies from 9 to 300. Also, the MBS-OEA and several well-designed existing algorithms are compared. The results show that the MBS-OEA can find high quality solutions for various problems. Additionally, the MBS-OEA shows a good performance in solving the problems with 300 hard rectangular blocks, 100 soft rectangular blocks, and 100 hybrid blocks, including both soft rectangular and hard rectilinear blocks. This illustrates that the MBS-OEA is not only suitable for solving a wide range of problems, but also competent for solving large-scale problems. Finally, a set of specific experiments is designed to identify the key component that is mainly responsible for the good performance of the MBS-OEA.

Taking inspiration from the interacting process among organizations in human societies, this correspondence designs a kind of structured population and corresponding evolutionary operators to form a novel algorithm, Organizational Evolutionary Algorithm (OEA), for solving both unconstrained and constrained optimization problems. In OEA, a population consists of organizations, and an organization consists of individuals. All evolutionary operators are designed to simulate the interaction among organizations. In experiments, 15 unconstrained functions, 13 constrained functions, and 4 engineering design problems are used to validate the performance of OEA, and thorough comparisons are made between the OEA and the existing approaches. The results show that the OEA obtains good performances in both the solution quality and the computational cost. Moreover, for the constrained problems, the good performances are obtained by only incorporating two simple constraints handling techniques into the OEA. Furthermore, systematic analyses have been made on all parameters of the OEA. The results show that the OEA is quite robust and easy to use.

Taking inspiration from the interacting process among organizations in human societies, a new classification algorithm, organizational coevolutionary algorithm for classification (OCEC), is proposed with the intrinsic properties of classification in mind. The main difference between OCEC and the available classification approaches based on evolutionary algorithms (EAs) is its use of a bottom-up search mechanism. OCEC causes the evolution of sets of examples, and at the end of the evolutionary process, extracts rules from these sets. These sets of examples form organizations. Because organizations are different from the individuals in traditional EAs, three evolutionary operators and a selection mechanism are devised for realizing the evolutionary operations performed on organizations. This method can avoid generating meaningless rules during the evolutionary process. An evolutionary method is also devised for determining the significance of each attribute, on the basis of which, the fitness function for organizations is defined. In experiments, the effectiveness of OCEC is first evaluated by multiplexer problems. Then, OCEC is compared with several well-known classification algorithms on 12 benchmarks from the UCI repository datasets and multiplexer problems. Moreover, OCEC is applied to a practical case, radar target recognition problems. All results show that OCEC achieves a higher predictive accuracy and a lower computational cost. Finally, the scalability of OCEC is studied on synthetic datasets. The number of training examples increases from 100 000 to 10 million, and the number of attributes increases from 9 to 400. The results show that OCEC obtains a good scalability.

Since all graphs in the 1993 DIMACS graph coloring challenge are undirected, each edge should be only counted once. However, in some files each edge is counted once, whereas in others each edge is counted twice; so a systematical check on the DIMACS challenge is made to eliminate the inconsistencies. Besides, the experimental results of a previous paper by Blas et al. counted each violated edges twice and neglected the inconsistencies in the DIMACS challenge. So the correct experimental results of a previous paper by Blas et al are also given.

With the intrinsic properties of constraint satisfaction problems (CSPs) in mind, we divide CSPs into two types, namely, permutation CSPs and nonpermutation CSPs. According to their characteristics, several behaviors are designed for agents by making use of the ability of agents to sense and act on the environment. These behaviors are controlled by means of evolution, so that the multiagent evolutionary algorithm for constraint satisfaction problems (MAEA-CSPs) results. To overcome the disadvantages of the general encoding methods, the minimum conflict encoding is also proposed. Theoretical analyzes show that MAEA-CSPs has a linear space complexity and converges to the global optimum. The first part of the experiments uses 250 benchmark binary CSPs and 79 graph coloring problems from the DIMACS challenge to test the performance of MAEA-CSPs for nonpermutation CSPs. MAEA-CSPs is compared with six well-defined algorithms and the effect of the parameters is analyzed systematically. The second part of the experiments uses a classical CSP, n-queen problems, and a more practical case, job-shop scheduling problems (JSPs), to test the performance of MAEA-CSPs for permutation CSPs. The scalability of MAEA-CSPs along n for n-queen problems is studied with great care. The results show that MAEA-CSPs achieves good performance when n increases from 10(4) to 10(7), and has a linear time complexity. Even for 10(7)-queen problems, MAEA-CSPs finds the solutions by only 150 seconds. For JSPs, 59 benchmark problems are used, and good performance is also obtained.

In this paper, multiagent systems and genetic algorithms are integrated to form a new algorithm, multiagent genetic algorithm (MAGA), for solving the global numerical optimization problem. An agent in MAGA represents a candidate solution to the optimization problem in hand. All agents live in a latticelike environment, with each agent fixed on a lattice-point. In order to increase energies, they compete or cooperate with their neighbors, and they can also use knowledge. Making use of these agent-agent interactions, MAGA realizes the purpose of minimizing the objective function value. Theoretical analyzes show that MAGA converges to the global optimum. In the first part of the experiments, ten benchmark functions are used to test the performance of MAGA, and the scalability of MAGA along the problem dimension is studied with great care. The results show that MAGA achieves a good performance when the dimensions are increased from 20-10,000. Moreover, even when the dimensions are increased to as high as 10,000, MAGA still can find high quality solutions at a low computational cost. Therefore, MAGA has good scalability and is a competent algorithm for solving high dimensional optimization problems. To the best of our knowledge, no researchers have ever optimized the functions with 10,000 dimensions by means of evolution. In the second part of the experiments, MAGA is applied to a practical case, the approximation of linear systems, with a satisfactory result.