A Multiscale Point-Supervised Network for Counting Maize Tassels in the Wild.

Accurate counting of maize tassels is essential for monitoring crop growth and estimating crop yield. Recently, deep-learning-based object detection methods have been used for this purpose, where plant counts are estimated from the number of bounding boxes detected. However, these methods suffer from 2 issues: (a) The scales of maize tassels vary because of image capture from varying distances and crop growth stage; and (b) tassel areas tend to be affected by occlusions or complex backgrounds, making the detection inefficient. In this paper, we propose a multiscale lite attention enhancement network (MLAENet) that uses only point-level annotations (i.e., objects labeled with points) to count maize tassels in the wild. Specifically, the proposed method includes a new multicolumn lite feature extraction module that generates a scale-dependent density map by exploiting multiple dilated convolutions with different rates, capturing rich contextual information at different scales more effectively. In addition, a multifeature enhancement module that integrates an attention strategy is proposed to enable the model to distinguish between tassel areas and their complex backgrounds. Finally, a new up-sampling module, UP-Block, is designed to improve the quality of the estimated density map by automatically suppressing the gridding effect during the up-sampling process. Extensive experiments on 2 publicly available tassel-counting datasets, maize tassels counting and maize tassels counting from unmanned aerial vehicle, demonstrate that the proposed MLAENet achieves marked advantages in counting accuracy and inference speed compared to state-of-the-art methods. The model is publicly available at https://github.com/ShiratsuyuShigure/MLAENet-pytorch/tree/main.

From Laboratory to Field: Unsupervised Domain Adaptation for Plant Disease Recognition in the Wild.

Plant disease recognition is of vital importance to monitor plant development and predicting crop production. However, due to data degradation caused by different conditions of image acquisition, e.g., laboratory vs. field environment, machine learning-based recognition models generated within a specific dataset (source domain) tend to lose their validity when generalized to a novel dataset (target domain). To this end, domain adaptation methods can be leveraged for the recognition by learning invariant representations across domains. In this paper, we aim at addressing the issues of domain shift existing in plant disease recognition and propose a novel unsupervised domain adaptation method via uncertainty regularization, namely, Multi-Representation Subdomain Adaptation Network with Uncertainty Regularization for Cross-Species Plant Disease Classification (MSUN). Our simple but effective MSUN makes a breakthrough in plant disease recognition in the wild by using a large amount of unlabeled data and via nonadversarial training. Specifically, MSUN comprises multirepresentation, subdomain adaptation modules and auxiliary uncertainty regularization. The multirepresentation module enables MSUN to learn the overall structure of features and also focus on capturing more details by using the multiple representations of the source domain. This effectively alleviates the problem of large interdomain discrepancy. Subdomain adaptation is used to capture discriminative properties by addressing the issue of higher interclass similarity and lower intraclass variation. Finally, the auxiliary uncertainty regularization effectively suppresses the uncertainty problem due to domain transfer. MSUN was experimentally validated to achieve optimal results on the PlantDoc, Plant-Pathology, Corn-Leaf-Diseases, and Tomato-Leaf-Diseases datasets, with accuracies of 56.06%, 72.31%, 96.78%, and 50.58%, respectively, surpassing other state-of-the-art domain adaptation techniques considerably.

A Segmentation-Guided Deep Learning Framework for Leaf Counting.

Deep learning-based methods have recently provided a means to rapidly and effectively extract various plant traits due to their powerful ability to depict a plant image across a variety of species and growth conditions. In this study, we focus on dealing with two fundamental tasks in plant phenotyping, i.e., plant segmentation and leaf counting, and propose a two-steam deep learning framework for segmenting plants and counting leaves with various size and shape from two-dimensional plant images. In the first stream, a multi-scale segmentation model using spatial pyramid is developed to extract leaves with different size and shape, where the fine-grained details of leaves are captured using deep feature extractor. In the second stream, a regression counting model is proposed to estimate the number of leaves without any pre-detection, where an auxiliary binary mask from segmentation stream is introduced to enhance the counting performance by effectively alleviating the influence of complex background. Extensive pot experiments are conducted CVPPP 2017 Leaf Counting Challenge dataset, which contains images of Arabidopsis and tobacco plants. The experimental results demonstrate that the proposed framework achieves a promising performance both in plant segmentation and leaf counting, providing a reference for the automatic analysis of plant phenotypes.

FPGA-based systolic deconvolution architecture for upsampling.

A deconvolution accelerator is proposed to upsample n × n input to 2n × 2n output by convolving with a k × k kernel. Its architecture avoids the need for insertion and padding of zeros and thus eliminates the redundant computations to achieve high resource efficiency with reduced number of multipliers and adders. The architecture is systolic and governed by a reference clock, enabling the sequential placement of the module to represent a pipelined decoder framework. The proposed accelerator is implemented on a Xilinx XC7Z020 platform, and achieves a performance of 3.641 giga operations per second (GOPS) with resource efficiency of 0.135 GOPS/DSP for upsampling 32 × 32 input to 256 × 256 output using a 3 × 3 kernel at 200 MHz. Furthermore, its high peak signal to noise ratio of almost 80 dB illustrates that the upsampled outputs of the bit truncated accelerator are comparable to IEEE double precision results.

Gait Recognition and Understanding Based on Hierarchical Temporal Memory Using 3D Gait Semantic Folding.

Gait recognition and understanding systems have shown a wide-ranging application prospect. However, their use of unstructured data from image and video has affected their performance, e.g., they are easily influenced by multi-views, occlusion, clothes, and object carrying conditions. This paper addresses these problems using a realistic 3-dimensional (3D) human structural data and sequential pattern learning framework with top-down attention modulating mechanism based on Hierarchical Temporal Memory (HTM). First, an accurate 2-dimensional (2D) to 3D human body pose and shape semantic parameters estimation method is proposed, which exploits the advantages of an instance-level body parsing model and a virtual dressing method. Second, by using gait semantic folding, the estimated body parameters are encoded using a sparse 2D matrix to construct the structural gait semantic image. In order to achieve time-based gait recognition, an HTM Network is constructed to obtain the sequence-level gait sparse distribution representations (SL-GSDRs). A top-down attention mechanism is introduced to deal with various conditions including multi-views by refining the SL-GSDRs, according to prior knowledge. The proposed gait learning model not only aids gait recognition tasks to overcome the difficulties in real application scenarios but also provides the structured gait semantic images for visual cognition. Experimental analyses on CMU MoBo, CASIA B, TUM-IITKGP, and KY4D datasets show a significant performance gain in terms of accuracy and robustness.

Shadow Detection in Still Road Images Using Chrominance Properties of Shadows and Spectral Power Distribution of the Illumination.

A well-known challenge in vision-based driver assistance systems is cast shadows on the road, which makes fundamental tasks such as road and lane detections difficult. In as much as shadow detection relies on shadow features, in this paper, we propose a set of new chrominance properties of shadows based on the skylight and sunlight contributions to the road surface chromaticity. Six constraints on shadow and non-shadowed regions are derived from these properties. The chrominance properties and the associated constraints are used as shadow features in an effective shadow detection method intended to be integrated on an onboard road detection system where the identification of cast shadows on the road is a determinant stage. Onboard systems deal with still outdoor images; thus, the approach focuses on distinguishing shadow boundaries from material changes by considering two illumination sources: sky and sun. A non-shadowed road region is illuminated by both skylight and sunlight, whereas a shadowed one is illuminated by skylight only; thus, their chromaticity varies. The shadow edge detection strategy consists of the identification of image edges separating shadowed and non-shadowed road regions. The classification is achieved by verifying whether the pixel chrominance values of regions on both sides of the image edges satisfy the six constraints. Experiments on real traffic scenes demonstrated the effectiveness of our shadow detection system in detecting shadow edges on the road and material-change edges, outperforming previous shadow detection methods based on physical features, and showing the high potential of the new chrominance properties.

Shadow-Based Vehicle Detection in Urban Traffic.

Vehicle detection is a fundamental task in Forward Collision Avoiding Systems (FACS). Generally, vision-based vehicle detection methods consist of two stages: hypotheses generation and hypotheses verification. In this paper, we focus on the former, presenting a feature-based method for on-road vehicle detection in urban traffic. Hypotheses for vehicle candidates are generated according to the shadow under the vehicles by comparing pixel properties across the vertical intensity gradients caused by shadows on the road, and followed by intensity thresholding and morphological discrimination. Unlike methods that identify the shadow under a vehicle as a road region with intensity smaller than a coarse lower bound of the intensity for road, the thresholding strategy we propose determines a coarse upper bound of the intensity for shadow which reduces false positives rates. The experimental results are promising in terms of detection performance and robustness in day time under different weather conditions and cluttered scenarios to enable validation for the first stage of a complete FACS.

Robust Arbitrary-View Gait Recognition Based on 3D Partial Similarity Matching.

Existing view-invariant gait recognition methods encounter difficulties due to limited number of available gait views and varying conditions during training. This paper proposes gait partial similarity matching that assumes a 3D object shares common view surfaces in significantly different views. Detecting such surfaces aids the extraction of gait features from multiple views; 3D parametric body models are morphed by pose and shape deformation from a template model using 2D gait silhouette as observation. The gait pose is estimated by a level set energy cost function from silhouettes including incomplete ones. Body shape deformation is achieved via Laplacian deformation energy function associated with inpainting gait silhouettes. Partial gait silhouettes in different views are extracted by gait partial region of interest elements selection and re-projected onto 2D space to construct partial gait energy images. A synthetic database with destination views and multi-linear subspace classifier fused with majority voting is used to achieve arbitrary view gait recognition that is robust to varying conditions. Experimental results on CMU, CASIA B, TUM-IITKGP, AVAMVG, and KY4D data sets show the efficacy of the propose method.

Significant body point labeling and tracking.

In this paper, a method is presented to label and track anatomical landmarks (e.g., head, hand/arm, feet), which are referred to as significant body points (SBPs), using implicit body models. By considering the human body as an inverted pendulum model, ellipse fitting and contour moments are applied to classify it as being in Stand, Sit, or Lie posture. A convex hull of the silhouette contour is used to determine the locations of SBPs. The particle filter or a motion flow-based method is used to predict SBPs in occlusion. Stick figures of various activities are generated by connecting the SBPs. The qualitative and quantitative evaluation show that the proposed method robustly labels and tracks SBPs in various activities of two different (low and high) resolution data sets.

2.5D multi-view gait recognition based on point cloud registration.

This paper presents a method for modeling a 2.5-dimensional (2.5D) human body and extracting the gait features for identifying the human subject. To achieve view-invariant gait recognition, a multi-view synthesizing method based on point cloud registration (MVSM) to generate multi-view training galleries is proposed. The concept of a density and curvature-based Color Gait Curvature Image is introduced to map 2.5D data onto a 2D space to enable data dimension reduction by discrete cosine transform and 2D principle component analysis. Gait recognition is achieved via a 2.5D view-invariant gait recognition method based on point cloud registration. Experimental results on the in-house database captured by a Microsoft Kinect camera show a significant performance gain when using MVSM.

Adaptive deformation correction of depth from defocus for object reconstruction.

The accuracy of three-dimensional object reconstruction using depth from defocus (DfD) can be severely reduced by elliptical lens deformation. This paper presents two correction methods, correction by deformation cancellation (CDC) and correction by least squares fit (CLSF). CDC works by subtracting the current deformed depth value by a prestored deformed value, and CLSF by mapping the deformed values to the expected values. Each method is followed by a smoothing algorithm to address the low-texture problem of DfD. Experiments using four DfD methods on real images show that the proposed methods effectively and efficiently eliminate the deformation.

Rational-operator-based depth-from-defocus approach to scene reconstruction.

This paper presents a rational-operator-based approach to depth from defocus (DfD) for the reconstruction of three-dimensional scenes from two-dimensional images, which enables fast DfD computation that is independent of scene textures. Two variants of the approach, one using the Gaussian rational operators (ROs) that are based on the Gaussian point spread function (PSF) and the second based on the generalized Gaussian PSF, are considered. A novel DfD correction method is also presented to further improve the performance of the approach. Experimental results are considered for real scenes and show that both approaches outperform existing RO-based methods.

Automatic image equalization and contrast enhancement using Gaussian mixture modeling.

In this paper, we propose an adaptive image equalization algorithm that automatically enhances the contrast in an input image. The algorithm uses the Gaussian mixture model to model the image gray-level distribution, and the intersection points of the Gaussian components in the model are used to partition the dynamic range of the image into input gray-level intervals. The contrast equalized image is generated by transforming the pixels' gray levels in each input interval to the appropriate output gray-level interval according to the dominant Gaussian component and the cumulative distribution function of the input interval. To take account of the hypothesis that homogeneous regions in the image represent homogeneous silences (or set of Gaussian components) in the image histogram, the Gaussian components with small variances are weighted with smaller values than the Gaussian components with larger variances, and the gray-level distribution is also used to weight the components in the mapping of the input interval to the output interval. Experimental results show that the proposed algorithm produces better or comparable enhanced images than several state-of-the-art algorithms. Unlike the other algorithms, the proposed algorithm is free of parameter setting for a given dynamic range of the enhanced image and can be applied to a wide range of image types.

Contextual and variational contrast enhancement.

This paper proposes an algorithm that enhances the contrast of an input image using interpixel contextual information. The algorithm uses a 2-D histogram of the input image constructed using a mutual relationship between each pixel and its neighboring pixels. A smooth 2-D target histogram is obtained by minimizing the sum of Frobenius norms of the differences from the input histogram and the uniformly distributed histogram. The enhancement is achieved by mapping the diagonal elements of the input histogram to the diagonal elements of the target histogram. Experimental results show that the algorithm produces better or comparable enhanced images than four state-of-the-art algorithms.

Local hull-based surface construction of volumetric data from silhouettes.

The marching cubes (MC) is a general method which can construct a surface of an object from its volumetric data generated using a shape from silhouette method. Although MC is efficient and straightforward to implement, a MC surface may have discontinuity even though the volumetric data is continuous. This is because surface construction is more sensitive to image noise than the construction of volumetric data. To address this problem, we propose a surface construction algorithm which aggregates local surfaces constructed by the 3-D convex hull algorithm. Thus, the proposed method initially classifies local convexities from imperfect MC vertices based on sliced volumetric data. Experimental results show that continuous surfaces are obtained from imperfect silhouette images of both convex and nonconvex objects.

Adaptive scale fixing for multiscale texture segmentation.

This paper addresses two challenging issues in unsupervised multiscale texture segmentation: determining adequate spatial and feature resolutions for different regions of the image, and utilizing information across different scales/resolutions. The center of a homogeneous texture is analyzed using coarse spatial resolution, and its border is detected using fine spatial resolution so as to locate the boundary accurately. The extraction of texture features is achieved via a multiresolution pyramid. The feature values are integrated across scales/resolutions adaptively. The number of textures is determined automatically using the variance ratio criterion. Experimental results on synthetic and real images demonstrate the improvement in performance of the proposed multiscale scheme over single scale approaches.

Efficient three-dimensional metric object modeling from uncalibrated image sequences.

This paper presents a scheme that addresses the practical issues associated with producing a geometric model of a scene using a passive sensing technique. The proposed image-based scheme comprises a recursive structure recovery method and a recursive surface reconstruction technique. The former method employs a robust corner-tracking algorithm that copes with the appearance and disappearance of features and a corner-based structure and motion estimation algorithm that handles the inclusion and expiration of features. The novel formulation and dual extended Kalman filter computational framework of the estimation algorithm provide an efficient approach to metric structure recovery that does not require any prior knowledge about the camera or scene. The newly developed surface reconstruction technique employs a visibility constraint to iteratively refine and ultimately yield a triangulated surface that envelops the recovered scene structure and can produce views consistent with those of the original image sequence. Results on simulated data and synthetic and real imagery illustrate that the proposed scheme is robust, accurate, and has good numerical stability, even when features are repeatedly absent or their image locations are affected by extreme levels of noise.

Image texture analysis of sonograms in chronic inflammations of thyroid gland.

The current practice in assessing sonographic findings of chronic inflamed thyroid tissue is mainly qualitative, based just on a physician's experience. This study shows that inflamed and healthy tissues can be differentiated by automatic texture analysis of B-mode sonographic images. Feature selection is the most important part of this procedure. We employed two selection schemes for finding recognition-optimal features: one based on compactness and separability and the other based on classification error. The full feature set included Muzzolini's spatial features and Haralick's co-occurrence features. These features were selected on a set of 2430 sonograms of 81 subjects, and the classifier performance was evaluated on a test set of 540 sonograms of 18 independent subjects. A classification success rate of 100% was achieved with as few as one optimal feature among the 129 texture characteristics tested. Both selection schemes agreed on the best features. The results were confirmed on the independent test set. The stability of the results with respect to sonograph setting, thyroid gland segmentation and scanning direction was tested.