Tens Of Pixels Research Articles

SuperYOLO: Super Resolution Assisted Object Detection in Multimodal Remote Sensing Imagery

Accurately and timely detecting multiscale small objects that contain tens of pixels from remote sensing images (RSI) remains challenging. Most of the existing solutions primarily design complex deep neural networks to learn strong feature representations for objects separated from the background, which often results in a heavy computation burden. In this article, we propose an accurate yet fast object detection method for RSI, named SuperYOLO, which fuses multimodal data and performs high-resolution (HR) object detection on multiscale objects by utilizing the assisted super resolution (SR) learning and considering both the detection accuracy and computation cost. First, we utilize a symmetric compact multimodal fusion (MF) to extract supplementary information from various data for improving small object detection in RSI. Furthermore, we design a simple and flexible SR branch to learn HR feature representations that can discriminate small objects from vast backgrounds with low-resolution (LR) input, thus further improving the detection accuracy. Moreover, to avoid introducing additional computation, the SR branch is discarded in the inference stage, and the computation of the network model is reduced due to the LR input. Experimental results show that, on the widely used VEDAI RS dataset, SuperYOLO achieves an accuracy of 75.09% (in terms of mAP50 ), which is more than 10% higher than the SOTA large models, such as YOLOv5l, YOLOv5x, and RS designed YOLOrs. Meanwhile, the parameter size and GFLOPs of SuperYOLO are about 18 times and 3.8 times less than YOLOv5x. Our proposed model shows a favorable accuracy and speed tradeoff compared to the state-of-the-art models. The code will be open-sourced at https://github.com/icey-zhang/SuperYOLO.

A Lightweight CNN Model for Refining Moving Vehicle Detection From Satellite Videos

With the recent development of earth observation technology, the satellites have obtained the ability to capture city-scale videos, which enable potential applications in intelligent traffic management. Because of the broad field-of-view, the moving vehicles in satellite videos are usually composed of only tens of pixels, making it difficult to differentiate true objects from noise and other distractors. In addition, the edges of tall building tops are often mistakenly detected as moving vehicles because of the effects of motion parallax. This article proposed a terse framework that can effectively suppress false targets, achieving high precision and recall. The study involves three parts: 1) An adaptive filtering method is proposed to reduce noise, thus making the detection algorithm more reliable; 2) Several background subtraction models are tested, and the best one is chosen to produce the preliminary detection results at high recall but low accuracy; 3) A lightweight convolutional neural network (LCNN) is designed and trained on a small collection of samples, and then used to eliminate false targets. The experiments and evaluations demonstrate that our method can largely improve the precision at the expense of a slight reduction of recall.

STATISTICS FOR PATCH OBSERVATIONS

In the application of remote sensing it is common to investigate processes that generate patches of material. This is especially true when using categorical land cover or land use maps. Here we view some existing tools, landscape pattern indices (LPI), as non-parametric estimators of random closed sets (RACS). This RACS framework enables LPIs to be studied rigorously. A RACS is any random process that generates a closed set, which encompasses any processes that result in binary (two-class) land cover maps. RACS theory, and methods in the underlying field of stochastic geometry, are particularly well suited to high-resolution remote sensing where objects extend across tens of pixels, and the shapes and orientations of patches are symptomatic of underlying processes. For some LPI this field already contains variance information and border correction techniques. After introducing RACS theory we discuss the core area LPI in detail. It is closely related to the spherical contact distribution leading to conditional variants, a new version of contagion, variance information and multiple border-corrected estimators. We demonstrate some of these findings on high resolution tree canopy data.

STATISTICS FOR PATCH OBSERVATIONS

Abstract. In the application of remote sensing it is common to investigate processes that generate patches of material. This is especially true when using categorical land cover or land use maps. Here we view some existing tools, landscape pattern indices (LPI), as non-parametric estimators of random closed sets (RACS). This RACS framework enables LPIs to be studied rigorously. A RACS is any random process that generates a closed set, which encompasses any processes that result in binary (two-class) land cover maps. RACS theory, and methods in the underlying field of stochastic geometry, are particularly well suited to high-resolution remote sensing where objects extend across tens of pixels, and the shapes and orientations of patches are symptomatic of underlying processes. For some LPI this field already contains variance information and border correction techniques. After introducing RACS theory we discuss the core area LPI in detail. It is closely related to the spherical contact distribution leading to conditional variants, a new version of contagion, variance information and multiple border-corrected estimators. We demonstrate some of these findings on high resolution tree canopy data.

Estimation of the characteristics of a promising orbital lidar for space debris monitoring

A preliminary study aimed at estimating the potential of a promising orbital lidar for space debris monitoring has been performed. Estimates of the space debris detection range under different observation conditions with different technical characteristics of lidar have been calculated. It is shown that the detection range for space debris fragments with sizes of 1 to 10 cm in the wavelength range near 0.5 µm depends strongly on the observation conditions, which are determined by the additive-noise level. It is concluded that a promising orbital lidar should contain a laser emitter with a power of several hundreds of watts and a receiving telescope about 0.5 to 1m in diameter. The possibility of using an orbital lidar for recognizing space debris is estimated. It is shown that a recognition probability exceeding 0.8 can be attained for a space debris image containing several tens of pixels and a signal-to-noise ratio above 10.

Integration of Chang'E-2 imagery and LRO laser altimeter data with a combined block adjustment for precision lunar topographic modeling

Lunar topographic information is essential for lunar scientific investigations and exploration missions. Lunar orbiter imagery and laser altimeter data are two major data sources for lunar topographic modeling. Most previous studies have processed the imagery and laser altimeter data separately for lunar topographic modeling, and there are usually inconsistencies between the derived lunar topographic models. This paper presents a novel combined block adjustment approach to integrate multiple strips of the Chinese Chang'E-2 imagery and NASA's Lunar Reconnaissance Orbiter (LRO) Laser Altimeter (LOLA) data for precision lunar topographic modeling. The participants of the combined block adjustment include the orientation parameters of the Chang'E-2 images, the intra-strip tie points derived from the Chang'E-2 stereo images of the same orbit, the inter-strip tie points derived from the overlapping area of two neighbor Chang'E-2 image strips, and the LOLA points. Two constraints are incorporated into the combined block adjustment including a local surface constraint and an orbit height constraint, which are specifically designed to remedy the large inconsistencies between the Chang'E-2 and LOLA data sets. The output of the combined block adjustment is the improved orientation parameters of the Chang'E-2 images and ground coordinates of the LOLA points, from which precision lunar topographic models can be generated. The performance of the developed approach was evaluated using the Chang'E-2 imagery and LOLA data in the Sinus Iridum area and the Apollo 15 landing area. The experimental results revealed that the mean absolute image residuals between the Chang'E-2 image strips were drastically reduced from tens of pixels before the adjustment to sub-pixel level after adjustment. Digital elevation models (DEMs) with 20 m resolution were generated using the Chang'E-2 imagery after the combined block adjustment. Comparison of the Chang'E-2 DEM with the LOLA DEM showed a good level of consistency. The developed combined block adjustment approach is of significance for the full comparative and synergistic use of lunar topographic data sets from different sensors and different missions.

High-quality curve rendering using line sampled visibility

Computing accurate visibility for thin primitives, such as hair strands, fur, grass, at all scales remains difficult or expensive. To that end, we present an efficient visibility algorithm based on spatial line sampling, and a novel intersection algorithm between line sample planes and Bézier splines with varying thickness. Our algorithm produces accurate visibility both when the projected width of the curve is a tiny fraction of a pixel, and when the projected width is tens of pixels. In addition, we present a rapid resolve procedure that computes final visibility. Using an optimized implementation running on graphics processors, we can render tens of thousands long hair strands with noise-free visibility at near-interactive rates.

Characterization of Barrett esophagus and esophageal adenocarcinoma by Fourier-transform infrared microscopy

The objective of this exploratory study was to evaluate the feasibility of using Fourier-Transform Infrared (FTIR) spectromicroscopy to characterize formalin-fixed, paraffin-embedded human esophageal tissues. Matched histologically normal esophageal squamous epithelium (NS), premalignant Barrett esophagus (BE), and primary esophageal adenocarcinoma (EADC) tissues, each defined according to strict clinicopathologic criteria, were obtained from patients who underwent esophageal resection. Using confocal IR microscopy, measurements in the mid-IR spectral region were carried out in transflection configuration, scanning regions of interest in 15 microm steps. A multidimensional dataset reporting the spectroscopic properties at each sampled point were analyzed by performing a hierarchical cluster analysis on the second derivative of spectral traces. Normal esophageal epithelia were characterized by a few well defined regions, mostly of large size (tens of contiguous pixels), which correlated with tissue histology, specifically the basal cell layer. BE tissues had characteristic regions localized to gland crypts, ranging in size from one pixel to a few tens of pixels, which displayed IR spectra with defined absorption features characteristic of glycoproteins. The incorporation of synchrotron light to improve the resolution of individual cells in BE tissues has demonstrated that these glycoproteins are associated with goblet cells, the characteristic cell type defining BE. Whereas the highly fragmented regions identified in EADC likely reflect tumor heterogeneity, FTIR mapping would appear to be a potentially useful technique to identify premalignant BE tissues. The technical feasibility of using FTIR to characterize formalin-fixed, paraffin-embedded human esophageal tissues demonstrates the potential of this technique to study archival human BE tissue specimens via automated screening techniques.

Mapping gossans in arid regions with Landsat TM and SIR-C images: the Beddaho Alteration Zone in northern Eritrea

Massive sulphide deposits in the Neoproterozoic Arabian-Nubian Shield are exposed at the surface as Fe-rich crusts termed gossans. Gossans are typically a few tens of metres across but are surrounded by wider clay- and Fe-rich alteration zones. Although Fe-rich gossans have characteristic reflectance spectra and surface roughness, they are often too small to be directly detected by Landsat TM or SIR-C images, both of which have about 30 m spatial resolution. In this paper, a procedure is described whereby gossans and the surrounding alteration zones can be identified and mapped by Landsat TM and SIR-C data using the Beddaho Alteration Zone and the Tebih Gossan in northern Eritrea as an example. Clay and Fe alteration index maps were generated by density slicing for Landsat TM band-ratios 5 7 and 3 1 , respectively. Landsat 5/7-4/5-3/1 TM images characteristically depict small (tens of pixels) gossans in blue and the more extensive alteration zones in pinkish purple. Chh-LhhLhh/Chh SIR-C images succeeded in identifying the gossan due to enhanced back-scattering of the radar shorter wavelength (6 cm) C-band by the rough gossan surfaces. This enhanced back-scattering might also be partially due to the characteristic dielectric property of the Fe-rich minerals forming the gossans. Choosing known gossans from both 5/7-4/5-3/1 Landsat TM and Chh-Lhh-Lhh/Chh SIR-C images as training sites for supervised classification helped to outline areas with reflectance spectra and radar back-scattering properties similar to those of the training sites. These results show significant correlation between supervised classifications based on the two data sets, suggesting a way to use combined visible and near infrared (VNIR) and radar imagery to explore for mineral deposits in arid regions.

Particle sizing from disk images by counting contiguous grid squares or vidicon pixels

The location of the light-scattering image of a particle with respect to the grid pattern of pixels on a vidicon screen used in particle surface monitoring can cause variation in the response of the vidicon for particles giving the same amount of light scattering. A simple disk-and-square model of the interaction is used to predict such geometrical interaction effects, which are shown to influence the mean number of pixels (squares) activated by particle images (disks) and the variation in the number of pixels activated. The same problem exists when trying to estimate disk area from squares covered or squares covered and touched. If one counts all squares touched by the disk, the mean number of squares, E( n), is shown to be exactly equal to the following equation, for image diameters ( D) smaller than the edge length ( D < L) and approximately equal to it for larger images: E(n) = 1 + 2( D L ) + ( π 4 )( D L ) 2 . The variance of n is calculated for D < L. The variation in n for D > L is demonstrated through geometrical modeling and square counting, distinguishing between the number of squares touched and the number completely covered. The geometrical interaction effects on the mean number of squares touched or mean number completely covered can be compensated for by calibration, but the effects on variation cannot. The variation of number of pixels activated compared to the mean number of pixels activated is predicted to be roughly 50% for situations where tens of pixels or less are activated by the particle image. Data that support this by showing substantially greater variability in the pixel distributions than the variability in the particle distributions are presented. The data show particles being missed in some readings and being counted in others, as the analysis predicts. The data also show relatively small error in the total counts, believed due to compensating errors.