Lateral Connections Research Articles

A neural network with competitive layers for character recognition

A structure and functioning mechanisms of a neural network with competitive layers are described. The network is intended to solve the character recognition task. The network consists of several competitive layers of neurons. Each layer is a neural network consisting of a number of neurons represented as a layer. The number of neural layers is equal to the number of recognized classes. All neural layers have one-to-one correspondence with one another and with the input raster. The neurons of every layer have mutual lateral learning connections, which weights are modified during the learning process. There is a competitive (inhibitory) relationship between all neural layers. This competitive interaction is realized by means of a “winner-take-all” (WTA) procedure which aim is to select the layer with the highest level of neural activity.Validation of the network has been done in experiments on recognition of handwritten digits of the MNIST database. The experiments have demonstrated that its error rate is few less than 2%, which is not a high result, but it is compensated by rather fast data processing and a very simple structure and functioning mechanisms.

Towards a more general understanding of the algorithmic utility of recurrent connections.

Lateral and recurrent connections are ubiquitous in biological neural circuits. Yet while the strong computational abilities of feedforward networks have been extensively studied, our understanding of the role and advantages of recurrent computations that might explain their prevalence remains an important open challenge. Foundational studies by Minsky and Roelfsema argued that computations that require propagation of global information for local computation to take place would particularly benefit from the sequential, parallel nature of processing in recurrent networks. Such "tag propagation" algorithms perform repeated, local propagation of information and were originally introduced in the context of detecting connectedness, a task that is challenging for feedforward networks. Here, we advance the understanding of the utility of lateral and recurrent computation by first performing a large-scale empirical study of neural architectures for the computation of connectedness to explore feedforward solutions more fully and establish robustly the importance of recurrent architectures. In addition, we highlight a tradeoff between computation time and performance and construct hybrid feedforward/recurrent models that perform well even in the presence of varying computational time limitations. We then generalize tag propagation architectures to propagating multiple interacting tags and demonstrate that these are efficient computational substrates for more general computations of connectedness by introducing and solving an abstracted biologically inspired decision-making task. Our work thus clarifies and expands the set of computational tasks that can be solved efficiently by recurrent computation, yielding hypotheses for structure in population activity that may be present in such tasks.

Experts in Action: Transnational Hong Kong–Style Stunt Work and Performance by Lauren Steimer

Reviewed by: Experts in Action: Transnational Hong Kong–Style Stunt Work and Performance by Lauren Steimer Karen Fang (bio) Experts in Action: Transnational Hong Kong–Style Stunt Work and Performance by Lauren Steimer. Duke University Press. 2021. 240 pages. $99.95 hardcover; $25.95 paper; also available in e-book. Chief among the many appeals of Quentin Tarantino's 2019 alternative history fable, Once Upon a Time in Hollywood, is the film's centering of a stuntman. "You're too pretty to be a stuntman," snarks the movie's cartoonishly bombastic Bruce Lee (Mike Moh) of the protagonist, played by Brad Pitt. In a typically Tarantino-ish mélange of film nerd and fan boy humor, the line reminds us that stunt workers, by definition, are never stars. By contrast, Lauren Steimer's new book, Experts in Action: Transnational Hong Kong–Style Stunt Work and Performance, is a scholarly exploration of the underlying premise that Tarantino (who is a recurring reference throughout her book) presents with such cheek. Situated between cinema and performance studies, and contributing to production culture, fan, and media industry studies, Steimer's work shines a spotlight on the "stunting stars" whose physical feats are a main attraction of contemporary action-centered film and television.1 Her specific interest is in how Hong Kong cinema's transnational influence has offered industrial and professional opportunities that might [End Page 207] not otherwise have been possible outside of Hollywood. Through case studies that include Thai action stars Tony Jaa and Jeeja Yanin, the international production of popular television series Xena: Warrior Princess (syndication, 1995–2001), and celebrated stunt doubles and action choreographers such as Dayna Grant, Chad Stahelski, and Zoë Bell, Steimer shows how global media and entertainment have assimilated Hong Kong production practices such as frontal performance, uncut action sequences, and a full-time second unit that enjoys a high degree of independent agency.2 These developments are part of the longer transnational flow of labor that has empowered Hollywood since the mid-1990s and whose effects have irrevocably transformed the look and feel of global media. Steimer analyzes the flexible, diffuse, truly global media landscape of the first decades of the twenty-first century, which scarcely resembles the unilateral, predominately transatlantic poaching of Anglo-European talent that enabled US global media dominance throughout the twentieth century. As Steimer puts it, adapting a phrase originally used by Xena director Doug Lefler to describe how he intentionally borrowed from Hong Kong action and production practice, Hong Kong is now a "reservoir of technique" whose impact on global media exceeds mainstream Hollywood cinema, spilling out from movies to television, documentary, and other content heavily influenced by fan fiction. This ripple effect is geographic as well as transmedial: while Hollywood lifts, in typically unabashed fashion, from Hong Kong, Hong Kong's global influence is further augmented by its Hollywood aura, fostering lateral connections with other non-Hollywood production sites—such as New Zealand—that benefit from their proximity to Hong Kong.3 Students and scholars of Hong Kong cinema, global media, performance, and industry studies will find much of value in Experts in Action. As a behind-the-scenes look into the specialized labor of contemporary stunts and physical performance, Steimer's book offers a fascinating glimpse into how the human spectacle of modern action cinema straddles both cutting-edge motion capture technologies and low-tech paraphernalia such as cardboard boxes to break falls. The book opens with a sustained discussion of practice theory and the scholarship on expertise, including Malcolm Gladwell's wellknown and oft-cited assertion of 10,000 hours as the minimum training time necessary to achieve mastery in any discipline. As example and model of such expertise, Steimer cites Jackie Chan's Peking opera training and Hong Kong stunt crew, whose influence and standard is evident throughout all the other performers and choreographers studied in Steimer's subsequent chapters.4 This paradigm of expert performance frames Steimer's overall goal of according attention and due credit to individual performers, action choreographers, and the "collective effort necessary to produce phenomenal body [End Page 208] effects."5 As Steimer wryly notes, "while, in the course of my research, I...

GHFormer-Net: Towards more accurate small green apple/begonia fruit detection in the nighttime

The accurate recognition of green fruit is of great significance for the automatic monitoring of the whole process of orchard growth, fruit harvesting, and yield estimation. However, accurate small green fruit detection, affected by many factors such as the unstructured orchard environment, the morphological diversity of target fruits, and multiple-scale fruits, is still an unsolved challenge, especially for small-scale fruits in the night environment. In this paper, we propose an optimized Retinanet-PVTv2 by introducing the gradient harmonizing mechanism to detect small green apple/begonia fruits in the night environment, namely GHFormer-Net. Specifically, PVTv2-B1 based on Transformer is applied as the backbone network to extract feature information from the global receptive, which breaks the limitation that spatial convolution is utilized to extract information from the local area; Next, with the help of FPN, shallow features and high-level features with rich semantic information are incorporated by lateral connections and a top-down structure to generate multi-scale feature maps; Then, a detector of RetinaNet is applied to detect green fruits. To adapt to small apple/begonia fruits detection in the night environment, a gradient harmonizing classification loss (GHM-R Loss) and a gradient harmonizing regression loss (GHM-R Loss) are introduced to improve RetinaNet-PVTv2. The experimental results show that our method achieves 85.2%/61.0% AP and 67.5%/45.2% APS on the benchmark of the NightFruit and Gala datasets respectively, which demonstrates the effectiveness of our method for small green apple/begonia fruits detection.

A spiking neural network model of the Superior Colliculus that is robust to changes in the spatial\u2013temporal input

Previous studies have indicated that the location of a large neural population in the Superior Colliculus (SC) motor map specifies the amplitude and direction of the saccadic eye-movement vector, while the saccade trajectory and velocity profile are encoded by the population firing rates. We recently proposed a simple spiking neural network model of the SC motor map, based on linear summation of individual spike effects of each recruited neuron, which accounts for many of the observed properties of SC cells in relation to the ensuing eye movement. However, in the model, the cortical input was kept invariant across different saccades. Electrical microstimulation and reversible lesion studies have demonstrated that the saccade properties are quite robust against large changes in supra-threshold SC activation, but that saccade amplitude and peak eye-velocity systematically decrease at low input strengths. These features were not accounted for by the linear spike-vector summation model. Here we show that the model’s input projection strengths and intra-collicular lateral connections can be tuned to generate saccades and neural spiking patterns that closely follow the experimental results.

Automatic concrete crack segmentation model based on transformer

Routine visual inspection of concrete structures is essential to maintain safe conditions. Therefore, studies of concrete crack segmentation using deep learning methods have been extensively conducted in recent years. However, insufficient performance remains a major challenge in diverse field-inspection scenarios. In this study, a novel SegCrack model for pixel-level crack segmentation is therefore proposed using a hierarchically structured Transformer encoder to output multiscale features and a top-down pathway with lateral connections to progressively up-sample and fuse features from the deepest layer of the encoder. Furthermore, an online hard example mining strategy was adopted to strengthen the detection of hard samples and improve the model performance. The effect of dataset size on the segmentation performance was then investigated. The results indicated that SegCrack achieved a precision, recall, F1 score, and mean intersection over union of 96.66%, 95.46%, 96.05%, and 92.63%, respectively, using the test set.

Improvement in Object Detection Using Multi-Scale RoI Pooling and Feature Pyramid Network

The feature pyramid network (FPN) enhances the localization accuracy and detection performance of small objects using multiple scales of the features. FPN adopts lateral connections and a top-down pathway to make low-level features semantically more meaningful. However, it uses only single-scale features to pool regions of interest (RoIs) when detecting objects. In this study, we showed that single-scale RoI pooling may not be the best solution for accurate localization and proposed multi-scale RoI pooling to improve the minor drawbacks of the FPN. The proposed method pools RoIs from three feature levels and concatenates the pooled features to detect objects. Thus, the FPN with multi-scale RoI pooling, called FPN+, detects objects by taking into account all information scattered across three feature levels. FPN+ improved the FPN by 2.81 and 1.1 points in COCO-style average precision (AP) when tested on PASCAL VOC 2007 test and COCO 2017 validation datasets, respectively.

Structural Connectivity of Human Inferior Colliculus Subdivisions Using in vivo and post mortem Diffusion MRI Tractography.

Inferior colliculus (IC) is an obligatory station along the ascending auditory pathway that also has a high degree of top-down convergence via efferent pathways, making it a major computational hub. Animal models have attributed critical roles for the IC in in mediating auditory plasticity, egocentric selection, and noise exclusion. IC contains multiple functionally distinct subdivisions. These include a central nucleus that predominantly receives ascending inputs and external and dorsal nuclei that receive more heterogeneous inputs, including descending and multisensory connections. Subdivisions of human IC have been challenging to identify and quantify using standard brain imaging techniques such as MRI, and the connectivity of each of these subnuclei has not been identified in the human brain. In this study, we estimated the connectivity of human IC subdivisions with diffusion MRI (dMRI) tractography, using both anatomical-based seed analysis as well as unsupervised k-means clustering. We demonstrate sensitivity of tractography to overall IC connections in both high resolution post mortem and in vivo datasets. k-Means clustering of the IC streamlines in both the post mortem and in vivo datasets generally segregated streamlines based on their terminus beyond IC, such as brainstem, thalamus, or contralateral IC. Using fine-grained anatomical segmentations of the major IC subdivisions, the post mortem dataset exhibited unique connectivity patterns from each subdivision, including commissural connections through dorsal IC and lateral lemniscal connections to central and external IC. The subdivisions were less distinct in the context of in vivo connectivity, although lateral lemniscal connections were again highest to central and external IC. Overall, the unsupervised and anatomically driven methods provide converging evidence for distinct connectivity profiles for each of the IC subdivisions in both post mortem and in vivo datasets, suggesting that dMRI tractography with high quality data is sensitive to neural pathways involved in auditory processing as well as top-down control of incoming auditory information.

Animal species detection and classification framework based on modified multi-scale attention mechanism and feature pyramid network

Detecting and classifying animal species is the first step in determining their long-term viability and the influence we may be having on them. Second, it aids people in recognizing predators and non-predatory animals, both of which pose a significant threat to humans and the environment. Third, it lowers the rate of traffic accidents in various regions since it has been a regular sighting on roadways, resulting in several collisions with automobiles. However, animal species' detection and Classification of animal species face many challenges such as the size and inconsistent behaviors various among the species. This paper proposes using a novel two-stage network with a modified multi-scale attention mechanism to create a more integrated recognition and classification system to attend to the challenges. At the regional proposal stage, a deeply characterized pyramid design with lateral connections was adopted, making the semantic characteristic of a small item more sensitive. Secondly, by reason of a densely connected convolutional network, the functional transmission is enhanced and multiplexed throughout the classification stage, resulting in a more precise Classification with fewer parameters. The Proposed model was evaluated using the AP and mAP evaluation metrics on the Animal wildlife and the challenging Animal-80 dataset. An mAP of +0.1% and an AP of 5% to 20% increase in each class was achieved by the attention-based proposed model compared to the non-attention-based model. Further comparison with other related works shows the proposed techniques' effectiveness for detecting and classifying animal species.

Predicting remaining useful life of turbofan engines using degradation signal based echo state network

Abstract The field of Prognostics and Health Management (PHM) in industries is gaining greater popularity to achieve high reliability by shifting the preventive maintenance to predictive maintenance. Estimation of Remaining Useful Life (RUL) is an effective prognostic measure that forecasts the health state of machine based on degradation modelling and condition monitoring. This article proposes a novel and robust methodology that uses Reduced Affinity Propagation (RAP) clustering technique that extracts representatives from the temporal signals measured through various heterogeneous sensors to predict the RUL using Echo State Network (ESN) with dynamic lateral inhibiting connections. The main advantage of the proposed model is that it does not overlook the features from the degradation signals and also learns the natural mapping among the representative points from the integrated sensor value. This approach is verified using CMAPPS dataset to show hopeful results in predicting the RUL of aircraft turbo fan engine. Also, this methodology can be a deployed as a tool in real time industrial applications to schedule predictive maintenance activities.

De Hemptinne, the Benedictines and Catholic Assimilation on the Congolese Copperbelt, 1911–1960

This article explores the history of the Benedictines in south-eastern Congo. The Benedictine leader, Jean-Félix De Hemptinne, eschewed an adaptationist approach to his mission work in favour of an assimilationist one. This article explains why he was able to follow such an approach for so long. Two factors were paramount. First, what Chris Bayly described as ‘lateral connections’ enabled De Hemptinne to side-step the need to engage meaningfully with local agricultural knowledge. Secondly, De Hemptinne's close if turbulent relationship with the colonial state facilitated a supply of funds and African labour despite the difficulties the Benedictines had in converting local people.

VC-YOLO: Towards Real-Time Object Detection in Aerial Images

Object detection for aerial images is a crucial and challenging task in the field of computer vision. Previous CNN-based methods face problems related to extreme variation of object scales and the complex background in aerial images, which vary significantly from natural scenes. On the other hand, a great many of existing detectors highly rely on computational performance and cannot handle real-time tasks. To address this problems, we propose a lightweight real-time object detection network which is named VC-YOLO. In the backbone part, we introduce a receptive field extended backbone with limited number of convolution layers to learn the features and context information of various objects. In the detection part, channel attention module and spatial attention module are used to generate discriminative feature representation. To make full use of semantic feature maps in backbone network, we improve the feature pyramid network (FPN) with more lateral connections to reuse the features in each convolution stage. We evaluate VC-YOLO on NWPU VHR-10 and VisDrone benchmark datasets. Experimental results show that VC-YOLO achieves superior detection accuracy with high efficiency compared with the existing methods.

“The boundary has been moved”: Hollywood cinéma-monde, film borders, and the multilingual assassin in Sicario and Inglourious Basterds

Transnational coproductions, multilingual dialogue, and border-crossing of many forms are growing increasingly common in contemporary cinemas. As a result, assigning a nationality to a film can prove a slippery and even arbitrary process. This article takes a new approach to films such as Sicario (Denis Villeneuve 2015) and Inglourious Basterds (Quentin Tarantino 2009), analyzing texts traditionally viewed as American through the lens of cinéma-monde (Marshall 2012). It focuses in particular on these films’ use of maps, and on their strikingly similar multilingual assassination scenes, reading them through Bill Marshall’s characterization of a cinema that “dramatically focuses attention on four elements: borders, movement, language, and lateral connections” (42). Each of these films was directed by an established auteur working in a “foreign” space and non-native languages, and each depicts continual border-crossing, code-switching, and violence committed across geographic and linguistic lines. With significant American and other characteristics, neither Sicario nor Inglourious Basterds could be neatly categorized as Quebecois nor French respectively. Yet these films implicate the French-speaking world in diverse ways. Ultimately, the ways in which these films traverse, theorize, and weaponize the border begs a questioning of how far the concepts of national cinemas, and indeed of cinéma-monde, can be extended.

SDNN: Symmetric deep neural networks with lateral connections for recommender systems

The recommender system is the key approach to alleviate the data explosion problem. Recently, with the rapid development of deep learning, there are several researches of employing deep neural networks (DNNs) on recommender systems. Most of these methods tend to capture the complex mapping relations between user-item representation and matching score via DNNs. These methods are mainly a pyramid structure which maps relations into low-dimensional space and then predicts the result by logistic regression. However, partial relations may be linearly indivisible in low-dimensional space. As we know, data that are hard to be separated in low-dimensional space can become much easier after being mapped into a high-dimensional space. Hence, motivated by the ladder network, we propose a Symmetric Deep Neural Networks (SDNN) with lateral connections, which can learn relations in both high-dimensional and low-dimensional spaces simultaneously. Moreover, considering that deep neural network is very inefficient in catching low-rank relations between users and items, we further combine SDNN with an improved deep matrix factorization model into a unified framework, and name this new model DualCF. Extensive experiments on three benchmark datasets are conducted and the results verify the effectiveness of SDNN and DualCF over state-of-the-art models for implicit feedback prediction.

Single Circuit in V1 Capable of Switching Contexts During Movement Using an Inhibitory Population as a Switch.

As animals adapt to their environments, their brains are tasked with processing stimuli in different sensory contexts. Whether these computations are context dependent or independent, they are all implemented in the same neural tissue. A crucial question is what neural architectures can respond flexibly to a range of stimulus conditions and switch between them. This is a particular case of flexible architecture that permits multiple related computations within a single circuit. Here, we address this question in the specific case of the visual system circuitry, focusing on context integration, defined as the integration of feedforward and surround information across visual space. We show that a biologically inspired microcircuit with multiple inhibitory cell types can switch between visual processing of the static context and the moving context. In our model, the VIP population acts as the switch and modulates the visual circuit through a disinhibitory motif. Moreover, the VIP population is efficient, requiring only a relatively small number of neurons to switch contexts. This circuit eliminates noise in videos by using appropriate lateral connections for contextual spatiotemporal surround modulation, having superior denoising performance compared to circuits where only one context is learned. Our findings shed light on a minimally complex architecture that is capable of switching between two naturalistic contexts using few switching units.

Construction of graphene network in Ni matrix composites: A molecular dynamics study of densification process

Graphene network (GN) reinforced metal matrix composites have been proved to possess a combination of high strength and good ductility, but the formation mechanism of GN, which serves as the source of excellent mechanical properties, is rarely revealed. In this study, the construction of network architecture in graphene network reinforced Ni matrix composites (GN/Ni) is investigated by molecular dynamics (MD). Results show that discrete graphene sheets covered on Ni particles are assembled into an intact and continuous graphene network through interlayer and lateral connections. The curvature of GN walls caused by defects, together with the corners of GN near the intersections of multi Ni particles, realizes the 3D network architecture composed of nondevelopable surfaces. The GN intertwines intimately with the Ni matrix to construct an interlocking structure, where a robust interface is induced by the strong Ni–C interactions and the redistribution of internal stress. Besides, owing to the restricted atomic motion of Ni matrix by GN during densification, dislocation accumulation near the GN is actualized in GN/Ni, as well as the grain refinement. This work can benefit the optimization of graphene reinforced metal matrix composites for fulfilling the strengthening and toughening potentials of GN.

A Neurodynamic Model of Saliency Prediction in V1.

Lateral connections in the primary visual cortex (V1) have long been hypothesized to be responsible for several visual processing mechanisms such as brightness induction, chromatic induction, visual discomfort, and bottom-up visual attention (also named saliency). Many computational models have been developed to independently predict these and other visual processes, but no computational model has been able to reproduce all of them simultaneously. In this work, we show that a biologically plausible computational model of lateral interactions of V1 is able to simultaneously predict saliency and all the aforementioned visual processes. Our model's architecture (NSWAM) is based on Penacchio's neurodynamic model of lateral connections of V1. It is defined as a network of firing rate neurons, sensitive to visual features such as brightness, color, orientation, and scale. We tested NSWAM saliency predictions using images from several eye tracking data sets. We show that the accuracy of predictions obtained by our architecture, using shuffled metrics, is similar to other state-of-the-art computational methods, particularly with synthetic images (CAT2000-Pattern and SID4VAM) that mainly contain low-level features. Moreover, we outperform other biologically inspired saliency models that are specifically designed to exclusively reproduce saliency. We show that our biologically plausible model of lateral connections can simultaneously explain different visual processes present in V1 (without applying any type of training or optimization and keeping the same parameterization for all the visual processes). This can be useful for the definition of a unified architecture of the primary visual cortex.

A connectivity-constrained computational account of topographic organization in primate high-level visual cortex

Inferotemporal (IT) cortex in humans and other primates is topographically organized, containing multiple hierarchically organized areas selective for particular domains, such as faces and scenes. This organization is commonly viewed in terms of evolved domain-specific visual mechanisms. Here, we develop an alternative, domain-general and developmental account of IT cortical organization. The account is instantiated in interactive topographic networks (ITNs), a class of computational models in which a hierarchy of model IT areas, subject to biologically plausible connectivity-based constraints, learns high-level visual representations optimized for multiple domains. We find that minimizing a wiring cost on spatially organized feedforward and lateral connections, alongside realistic constraints on the sign of neuronal connectivity within model IT, results in a hierarchical, topographic organization. This organization replicates a number of key properties of primate IT cortex, including the presence of domain-selective spatial clusters preferentially involved in the representation of faces, objects, and scenes; columnar responses across separate excitatory and inhibitory units; and generic spatial organization whereby the response correlation of pairs of units falls off with their distance. We thus argue that topographic domain selectivity is an emergent property of a visual system optimized to maximize behavioral performance under generic connectivity-based constraints.

A comparative approach to quantify the heterarchical structures of complex systems

The dynamics and adaptive capacity of social-ecological systems are heavily contingent on system structure, which is established through geography, institutions, interactions, and movement. Contrasting views of system structure, as hierarchies and single-level networks respectively, have tended to emphasize the role of either top-down or lateral (peer-to-peer) connections. The concept of a heterarchy aims to capture both top-down and lateral connections on orthogonal axes and has been proposed as a way of unifying alternative approaches to measuring structure, but it has not been fully operationalized for quantifying and comparing system structures. We developed a simple approach to consistently quantifying heterarchical structure across different kinds of networks. We first calculated suitable metrics, including modularity and a hierarchy score, for a wide range of both simulated and real-world systems including food webs, biological, infrastructure, and social networks. Metrics were corrected for differences in size and magnitude. The results were then visualized as a heterarchy matrix. We compared the angle (degrees) and Euclidian distance of each simulated and real-world network from the center of the matrix between network groups. All networks showed distinct placement on the heterarchy matrix. Relative to one another, food webs were laterally polycentric, social networks were mainly pyramidal and coordinated polycentric, and biological networks were pyramidal and laterally polycentric. Our test of concept, although relatively basic, provides strong evidence that system structure cannot be fully understood as purely laterally connected or purely hierarchical. System resilience requires a trade-off between modularity, aiding redundancy and collaboration; and hierarchy, aiding efficient action. Our approach has the potential to provide a robust, accessible methodology to quantify system structure that allows for universal contextualization, a key step within fields such as resilience and sustainability science.

Lateral Feature Enhancement Network for Page Object Detection

In this paper, lateral feature enhancement (LFE) backbone network is proposed to enrich feature representation effectively for page object detection across various scales. Our LFE backbone network has three feature enhancement modules. Firstly, feature enhancement of large page object is a bottom-up feature pyramid, enhancing features of large page objects which convey more important information to readers. Secondly, lateral feature enhancement includes a top-down feature pyramid propagating representative semantical features to lower layers, and a lateral connection for feature enhancement in each layer. Thirdly, lateral skip connection is designed to retain the original feature details. The stacking strategies of bottom-up, top-down and lateral connections are beneficial for overall object detection. Visualization of feature indicates that the proposed LFE backbone network enhances global semantic information as well as detailed features of small page objects. Comparative experiments on the two state-of-the-art datasets show that it achieves excellent results with 0.950 mAP on PubLayNet and 0.892 mAP on POD with more strict metric IoU=0.8 respectively. Compared with both computer vision (CV) based unimodal detectors and multi-modal detectors, the proposed LFE network performs excellently. Visual effect experiments compare performances of CV-based detectors. The results show our detector outperforms others with strict metric, especially in the detection of small page objects.