DuFeNet: Improve the Accuracy and Increase Shape Bias of Neural Network Models

In image classification field, existing work tends to modify the network structure to obtain higher accuracy or faster speed. However, some studies have found that the neural network usually has texture bias effect, which means that the neural network is more sensitive to the texture information than the shape information. Based on such phenomenon, we propose a new way to improve network performance by making full use of gradient information. The dual features network (DuFeNet) is proposed in this paper. In DuFeNet, one sub-network is used to learn the information of gradient features, and the other is a traditional neural network with texture bias. The structure of DuFeNet is easy to implement in the original neural network structure. The experimental results clearly show that DuFeNet can achieve better accuracy in image classification and detection. It can increase the shape bias of the network adapted to human visual perception. Besides, DuFeNet can be used without modifying the structure of the original network at lower additional parameters cost.

Few Samples of SAR Automatic Target Recognition Based on Enhanced-Shape CNN

Synthetic Aperture Radar (SAR), as one of the important and significant methods for obtaining target characteristics in the field of remote sensing, has been applied to many fields including intelligence search, topographic surveying, mapping, and geological survey. In SAR field, the SAR automatic target recognition (SAR ATR) is a significant issue. However, on the other hand, it also has high application value. The development of deep learning has enabled it to be applied to SAR ATR. Some researchers point out that existing convolutional neural network (CNN) paid more attention to texture information, which is often not as good as shape information. Wherefore, this study designs the enhanced-shape CNN, which enhances the target shape at the input. Further, it uses an improved attention module, so that the network can highlight target shape in SAR images. Aiming at the problem of the small scale of the existing SAR data set, a small sample experiment is conducted. Enhanced-shape CNN achieved a recognition rate of 99.29% when trained on the full training set, while it is 89.93% on the one-eighth training data set.

IR-MF-SVMe: Image Retrieval using Multiple Features Extraction with Support Vector Machine ensemble

Image representation is a widespread strategy of image retrieval based on appearance, shape information. The traditional feature representation methods ignore hidden information that exists in the dataset samples; it reduces the discriminative performance of the classifier and excludes various geometric and photometric variations consideration in obtaining the features; these degrade retrieval performance. Hence, proposed multiple features fusion and Support Vector Machines Ensemble (IR-MF-SVMe); an Image Retrieval framework to enhance the performance of the retrieval process. The Color Histogram (CH), Color Auto-Correlogram (CAC), Color Moments (CM), Gabor Wavelet (GW), and Wavelet Moments (WM) descriptors are used to extract multiple features that separate the element vectors of images in representation. The multi-class classifier is constructed with the aggregation of binary Support Vector Machines, which decrease the count of false positives within the interrelated semantic classes. The proposed framework is validated on the WANG dataset and results in the accuracy of 84% for the individual features and 86% for the fused features related to the state-of-the-arts.

Bringing Home Baby Euclid: Testing Infants’ Basic Shape Discrimination Online

Online developmental psychology studies are still in their infancy, but their role is newly urgent in the light of the COVID-19 pandemic and the suspension of in-person research. Are online studies with infants a suitable stand-in for laboratory-based studies? Across two unmonitored online experiments using a change-detection looking-time paradigm with 96 7-month-old infants, we found that infants did not exhibit measurable sensitivities to the basic shape information that distinguishes between 2D geometric forms, as had been observed in previous laboratory experiments. Moreover, while infants were distracted in our online experiments, such distraction was nevertheless not a reliable predictor of their ability to discriminate shape information. Our findings suggest that the change-detection paradigm may not elicit infants’ shape discrimination abilities when stimuli are presented on small, personal computer screens because infants may not perceive two discrete events with only one event displaying uniquely changing information that draws their attention. Some developmental paradigms used with infants, even those that seem well-suited to the constraints and goals of online data collection, may thus not yield results consistent with the laboratory results that rely on highly controlled settings and specialized equipment, such as large screens. As developmental researchers continue to adapt laboratory-based methods to online contexts, testing those methods online is a necessary first step in creating robust tools and expanding the space of inquiry for developmental science conducted online.

Animal Migration Patterns Extraction Based on Atrous-Gated CNN Deep Learning Model

Weather radar data can capture large-scale bird migration information, helping solve a series of migratory ecological problems. However, extracting and identifying bird information from weather radar data remains one of the challenges of radar aeroecology. In recent years, deep learning was applied to the field of radar data processing and proved to be an effective strategy. This paper describes a deep learning method for extracting biological target echoes from weather radar images. This model uses a two-stream CNN (Atrous-Gated CNN) architecture to generate fine-scale predictions by combining the key modules such as squeeze-and-excitation (SE), and atrous spatial pyramid pooling (ASPP). The SE block can enhance the attention on the feature map, while ASPP block can expand the receptive field, helping the network understand the global shape information. The experiments show that in the typical historical data of China next generation weather radar (CINRAD), the precision of the network in identifying biological targets reaches up to 99.6%. Our network can cope with complex weather conditions, realizing long-term and automated monitoring of weather radar data to extract biological target information and provide feasible technical support for bird migration research.

The Effect of Spatial Frequency on the Visual Category Representation in the Macaque Inferior Temporal Cortex

ABSTRACTTo expand our knowledge about the object recognition, it is critical to understand the role of spatial frequency (SF) in an object representation that occurs in the inferior temporal (IT) cortex at the final stage of processing the visual information across the ventral visual pathway. Object categories are being recognized hierarchically in at least three levels of abstraction: superordinate (e.g., animal), mid-level (e.g., human face), and subordinate (e.g., face identity). Psychophysical studies have shown rapid access to mid-level category information and low SF (LSF) contents. Although the hierarchical representation of categories has been shown to exist inside the IT cortex, the impact of SF on the multi-level category processing is poorly understood. To gain a deeper understanding of the neural basis of the interaction between SF and category representations at multiple levels, we examined the neural responses within the IT cortex of macaque monkeys viewing several SF-filtered objects. Each stimulus could be either intact or bandpass filtered into either the LSF (coarse shape information) or high SF (HSF) (fine shape information) bands. We found that in both High- and Low-SF contents, the advantage of mid-level representation has not been violated. This evidence suggests that mid-level category boundary maps are strongly represented in the IT cortex and remain unaffected with respect to any changes in the frequency content of stimuli. Our observations indicate the necessity of the HSF content for the superordinate category representation inside the IT cortex. In addition, our findings reveal that the representation of global category information is more dependent on the HSF than the LSF content. Furthermore, the lack of subordinate representation in both LSF and HSF filtered stimuli compared to the intact stimuli provide strong evidence that all SF contents are necessary for fine category visual processing.

A Graph Optimization-Based Acoustic SLAM Edge Computing System Offering Centimeter-Level Mapping Services with Reflector Recognition Capability

Robots can use echo signals for simultaneous localization and mapping (SLAM) services in unknown environments where its own camera is not available. In current acoustic SLAM solutions, the time of arrival (TOA) in the room impulse response (RIR) needs to be associated with the corresponding reflected wall, which leads to an echo labelling problem (ELP). The position of the wall can be derived from the TOA associated with the wall, but most of the current solutions ignore the effect of the cumulative error in the robot’s moving state measurement on the wall position estimation. In addition, the estimated room map contains only the shape information of the room and lacks position information such as the positions of doors and windows. To address the above problems, this paper proposes a graph optimization-based acoustic SLAM edge computing system offering centimeter-level mapping services with reflector recognition capability. In this paper, a robot equipped with a sound source and a four-channel microphone array travels around the room, and it can collect the room impulse response at different positions of the room and extract the RIR cepstrum feature from the room impulse response. The ELP is solved by using the RIR cepstrum to identify reflectors with different absorption coefficients. Then, the similarity of the RIR cepstrum vectors is used for closed-loop detection. Finally, this paper proposes a method to eliminate the cumulative error of robot movement by fusing IMU data and acoustic echo data using graph-optimized edge computation. The experiments show that the acoustic SLAM system in this paper can accurately estimate the trajectory of the robot and the position of doors, windows, and so on in the room map. The average self-localization error of the robot is 2.84 cm, and the mapping error is 4.86 cm, which meet the requirement of centimeter-level map service.

ShapeFit: extracting the power spectrum shape information in galaxy surveys beyond BAO and RSD

In the standard (classic) approach, galaxy clustering measurements from spectroscopic surveys are compressed into baryon acoustic oscillations and redshift space distortions measurements, which in turn can be compared to cosmological models. Recent works have shown that avoiding this intermediate step and fitting directly the full power spectrum signal (full modelling) leads to much tighter constraints on cosmological parameters. Here we show where this extra information is coming from and extend the classic approach with one additional effective parameter, such that it captures, effectively, the same amount of information as the full modelling approach, but in a model-independent way. We validate this new method (ShapeFit) on mock catalogs, and compare its performance to the full modelling approach finding both to deliver equivalent results. The ShapeFit extension of the classic approach promotes the standard analyses at the level of full modelling ones in terms of information content, with the advantages of i) being more model independent; ii) offering an understanding of the origin of the extra cosmological information; iii) allowing a robust control on the impact of observational systematics.

A Novel Framework for Modeling Weakly Lensing Shear Using Kinematics and Imaging at Moderate Redshift

Kinematic weak lensing describes the distortion of a galaxy’s projected velocity field due to lensing shear, an effect recently reported for the first time by Gurri et al. based on a sample of 18 galaxies at z ∼ 0.1. In this paper, we develop a new formalism that combines the shape information from imaging surveys with the kinematic information from resolved spectroscopy to better constrain the lensing distortion of source galaxies and to potentially address systematic errors that affect conventional weak-lensing analyses. Using a Bayesian forward model applied to mock galaxy observations, we model distortions in the source galaxy’s velocity field simultaneously with the apparent shear-induced offset between the kinematic and photometric major axes. We show that this combination dramatically reduces the statistical uncertainty on the inferred shear, yielding statistical error gains of a factor of 2–6 compared to kinematics alone. While we have not accounted for errors from intrinsic kinematic irregularities, our approach opens kinematic lensing studies to higher redshifts where resolved spectroscopy is more challenging. For example, we show that ground-based integral-field spectroscopy of background galaxies at z ∼ 0.7 can deliver gravitational shear measurements with signal-to-noise ratio of ∼1 per source galaxy at 1 arcminute separations from a galaxy cluster at z ∼ 0.3. This suggests that even modest samples observed with existing instruments could deliver improved galaxy cluster mass measurements and well-sampled probes of their halo mass profiles to large radii.