Filter pruning for convolutional neural networks in semantic image segmentation.

The remarkable performance of Convolutional Neural Networks (CNNs) has increased their use in real-time systems and devices with limited resources. Hence, compacting these networks while preserving accuracy has become necessary, leading to multiple compression methods. However, the majority require intensive iterative procedures and do not delve into the influence of the used data. To overcome these issues, this paper presents several contributions, framed in the context of explainable Artificial Intelligence (xAI): (a) two filter pruning methods for CNNs, which remove the less significant convolutional kernels; (b) a fine-tuning strategy to recover generalization; (c) a layer pruning approach for U-Net; and (d) an explanation of the relationship between performance and the used data. Filter and feature maps information are used in the pruning process: Principal Component Analysis (PCA) is combined with a next-convolution influence-metric, while the latter and the mean standard deviation are used in an importance score distribution-based method. The developed strategies are generic, and therefore applicable to different models. Experiments demonstrating their effectiveness are conducted over distinct CNNs and datasets, focusing mainly on semantic segmentation (using U-Net, DeepLabv3+, SegNet, and VGG-16 as highly representative models). Pruned U-Net on agricultural benchmarks achieves 98.7% parameters and 97.5% FLOPs drop, with a 0.35% gain in accuracy. DeepLabv3+ and SegNet on CamVid reach 46.5% and 72.4% parameters reduction and a 51.9% and 83.6% FLOPs drop respectively, with almost no decrease in accuracy. VGG-16 on CIFAR-10 obtains up to 86.5% parameter and 82.2% FLOPs decrease with a 0.78% accuracy gain.

Artificial Intelligence Techniques for Automatic Detection of Peri-implant Marginal Bone Remodeling in Intraoral Radiographs.

Peri-implantitis can cause marginal bone remodeling around implants. The aim is to develop an automatic image processing approach based on two artificial intelligence (AI) techniques in intraoral (periapical and bitewing) radiographs to assist dentists in determining bone loss. The first is a deep learning (DL) object-detector (YOLOv3) to roughly identify (no exact localization is required) two objects: prosthesis (crown) and implant (screw). The second is an image understanding-based (IU) process to fine-tune lines on screw edges and to identify significant points (intensity bone changes, intersections between screw and crown). Distances between these points are used to compute bone loss. A total of 2920 radiographs were used for training (50%) and testing (50%) the DL process. The mAP@0.5 metric is used for performance evaluation of DL considering periapical/bitewing and screws/crowns in upper and lower jaws, with scores ranging from 0.537 to 0.898 (sufficient because DL only needs an approximation). The IU performance is assessed with 50% of the testing radiographs through the t test statistical method, obtaining p values of 0.0106 (line fitting) and 0.0213 (significant point detection). The IU performance is satisfactory, as these values are in accordance with the statistical average/standard deviation in pixels for line fitting (2.75/1.01) and for significant point detection (2.63/1.28) according to the expert criteria of dentists, who establish the ground-truth lines and significant points. In conclusion, AI methods have good prospects for automatic bone loss detection in intraoral radiographs to assist dental specialists in diagnosing peri-implantitis.

Automatic computation of mandibular indices in dental panoramic radiographs for early osteoporosis detection.

AIM: A new automatic method for detecting specific points and lines (straight and curves) in dental panoramic radiographies (orthopantomographies) is proposed, where the human knowledge is mapped to the automatic system. The goal is to compute relevant mandibular indices (Mandibular Cortical Width, Panoramic Mandibular Index, Mandibular Ratio, Mandibular Cortical Index) in order to detect the thinning and deterioration of the mandibular bone. Data can be stored for posterior massive analysis. METHODS: Panoramic radiographies are intrinsically complex, including: artificial structures, unclear limits in bony structures, jawbones with irregular curvatures and intensity levels, irregular shapes and borders of the mental foramen, irregular teeth alignments or missing dental pieces. An intelligent sequence of linked imaging segmentation processes is proposed to cope with the above situations towards the design of the automatic segmentation, making the following contributions: (i) Fuzzy K-means classification for identifying artificial structures; (ii) adjust a tangent line to the lower border of the lower jawbone (lower cortex), based on texture analysis, grey scale dilation, binarization and labelling; (iii) identification of the mental foramen region and its centre, based on multi-thresholding, binarization, morphological operations and labelling; (iv) tracing a perpendicular line to the tangent passing through the centre of the mental foramen region and two parallel lines to the tangent, passing through borders on the mental foramen intersected by the perpendicular; (v) following the perpendicular line, a sweep is made moving up the tangent for detecting accumulation of binary points after applying adaptive filtering; (vi) detection of the lower mandible alveolar crest line based on the identification of inter-teeth gaps by saliency and interest points feature description. RESULTS: The performance of the proposed approach was quantitatively compared against the criteria of expert dentists, verifying also its validity with statistical studies based on the analysis of deterioration of bone structures with different levels of osteoporosis. All indices are computed inside two regions of interest, which tolerate flexibility in sizes and locations, making this process robust enough. CONCLUSIONS: The proposed approach provides an automatic procedure able to process with efficiency and reliability panoramic X-Ray images for early osteoporosis detection.

Structural correlates of apathy in Alzheimer's disease: a multimodal MRI study.

OBJECTIVE: Apathy is one of the most frequent symptoms of dementia, whose underlying neurobiology is not well understood. The objective was to analyze the correlations of apathy and its dimensions with gray and white matter damage in the brain of patients with advanced Alzheimer's disease (AD). METHODS: The setting of the study was at the Alzheimer Center Reina Sofía Foundation Research Unit. Participants include 37 nursing home patients with moderate to severe AD, 78.4% were women, and mean Standard Deviation (SD) age is 82.7 (5.8). Several measurements were taken: severe mini-mental state examination and Global Deterioration Scale for cognitive and functional status, Neuropsychiatric Inventory for behavioral problems, and Apathy In Dementia-Nursing Home Version Scale for apathy, including total score and subscores of emotional blunting, deficit of thinking, and cognitive inertia. 3T magnetic resonance imaging measures (voxel-based morphometry, fluid-attenuated inversion recovery, and diffusion tensor imaging) were also conducted. RESULTS: Moderate levels of apathy (mean Apathy In Dementia-Nursing Home Version Scale: 31.1 ± 18.5) were found. Bilateral damage to the corpus callosum and internal capsule was associated with apathy severity (cluster size 2435, p < 0.0005, family-wise error [FWE]-corrected). A smaller and more anteriorly located region of the right internal capsule and corpus callosum was associated with higher emotional blunting (cluster size 334, p < 0.0005, FWE-corrected). Ischemic damage in the right periventricular frontal region was associated with higher deficit of thinking (cluster size 3805, p < 0.005, FWE-corrected). CONCLUSIONS: Brain damage related to apathy may have different features in the advanced stages of AD and differs between the three apathy dimensions. Besides atrophy, brain connectivity and vascular lesions are relevant in the study of apathy, especially in the more severe stages of dementia. Further magnetic resonance imaging studies should include multimodal techniques. Copyright © 2016 John Wiley & Sons, Ltd.

New trends in robotics for agriculture: integration and assessment of a real fleet of robots.

Computer-based sensors and actuators such as global positioning systems, machine vision, and laser-based sensors have progressively been incorporated into mobile robots with the aim of configuring autonomous systems capable of shifting operator activities in agricultural tasks. However, the incorporation of many electronic systems into a robot impairs its reliability and increases its cost. Hardware minimization, as well as software minimization and ease of integration, is essential to obtain feasible robotic systems. A step forward in the application of automatic equipment in agriculture is the use of fleets of robots, in which a number of specialized robots collaborate to accomplish one or several agricultural tasks. This paper strives to develop a system architecture for both individual robots and robots working in fleets to improve reliability, decrease complexity and costs, and permit the integration of software from different developers. Several solutions are studied, from a fully distributed to a whole integrated architecture in which a central computer runs all processes. This work also studies diverse topologies for controlling fleets of robots and advances other prospective topologies. The architecture presented in this paper is being successfully applied in the RHEA fleet, which comprises three ground mobile units based on a commercial tractor chassis.

Integrating sensory/actuation systems in agricultural vehicles.

In recent years, there have been major advances in the development of new and more powerful perception systems for agriculture, such as computer-vision and global positioning systems. Due to these advances, the automation of agricultural tasks has received an important stimulus, especially in the area of selective weed control where high precision is essential for the proper use of resources and the implementation of more efficient treatments. Such autonomous agricultural systems incorporate and integrate perception systems for acquiring information from the environment, decision-making systems for interpreting and analyzing such information, and actuation systems that are responsible for performing the agricultural operations. These systems consist of different sensors, actuators, and computers that work synchronously in a specific architecture for the intended purpose. The main contribution of this paper is the selection, arrangement, integration, and synchronization of these systems to form a whole autonomous vehicle for agricultural applications. This type of vehicle has attracted growing interest, not only for researchers but also for manufacturers and farmers. The experimental results demonstrate the success and performance of the integrated system in guidance and weed control tasks in a maize field, indicating its utility and efficiency. The whole system is sufficiently flexible for use in other agricultural tasks with little effort and is another important contribution in the field of autonomous agricultural vehicles.

Grapheme-color synesthetes show peculiarities in their emotional brain: cortical and subcortical evidence from VBM analysis of 3D-T1 and DTI data.

Grapheme-color synesthesia is a neurological phenomenon in which viewing achromatic letters/numbers leads to automatic and involuntary color experiences. In this study, voxel-based morphometry analyses were performed on T1 images and fractional anisotropy measures to examine the whole brain in associator grapheme-color synesthetes. These analyses provide new evidence of variations in emotional areas (both at the cortical and subcortical levels), findings that help understand the emotional component as a relevant aspect of the synesthetic experience. Additionally, this study replicates previous findings in the left intraparietal sulcus and, for the first time, reports the existence of anatomical differences in subcortical gray nuclei of developmental grapheme-color synesthetes, providing a link between acquired and developmental synesthesia. This empirical evidence, which goes beyond modality-specific areas, could lead to a better understanding of grapheme-color synesthesia as well as of other modalities of the phenomenon.

Camera sensor arrangement for crop/weed detection accuracy in agronomic images.

In Precision Agriculture, images coming from camera-based sensors are commonly used for weed identification and crop line detection, either to apply specific treatments or for vehicle guidance purposes. Accuracy of identification and detection is an important issue to be addressed in image processing. There are two main types of parameters affecting the accuracy of the images, namely: (a) extrinsic, related to the sensor's positioning in the tractor; (b) intrinsic, related to the sensor specifications, such as CCD resolution, focal length or iris aperture, among others. Moreover, in agricultural applications, the uncontrolled illumination, existing in outdoor environments, is also an important factor affecting the image accuracy. This paper is exclusively focused on two main issues, always with the goal to achieve the highest image accuracy in Precision Agriculture applications, making the following two main contributions: (a) camera sensor arrangement, to adjust extrinsic parameters and (b) design of strategies for controlling the adverse illumination effects.

White blood cell segmentation by circle detection using electromagnetism-like optimization.

Medical imaging is a relevant field of application of image processing algorithms. In particular, the analysis of white blood cell (WBC) images has engaged researchers from fields of medicine and computer vision alike. Since WBCs can be approximated by a quasicircular form, a circular detector algorithm may be successfully applied. This paper presents an algorithm for the automatic detection of white blood cells embedded into complicated and cluttered smear images that considers the complete process as a circle detection problem. The approach is based on a nature-inspired technique called the electromagnetism-like optimization (EMO) algorithm which is a heuristic method that follows electromagnetism principles for solving complex optimization problems. The proposed approach uses an objective function which measures the resemblance of a candidate circle to an actual WBC. Guided by the values of such objective function, the set of encoded candidate circles are evolved by using EMO, so that they can fit into the actual blood cells contained in the edge map of the image. Experimental results from blood cell images with a varying range of complexity are included to validate the efficiency of the proposed technique regarding detection, robustness, and stability.

Evaluation of a change detection methodology by means of binary thresholding algorithms and informational fusion processes.

Landcover is subject to continuous changes on a wide variety of temporal and spatial scales. Those changes produce significant effects in human and natural activities. Maintaining an updated spatial database with the occurred changes allows a better monitoring of the Earth's resources and management of the environment. Change detection (CD) techniques using images from different sensors, such as satellite imagery, aerial photographs, etc., have proven to be suitable and secure data sources from which updated information can be extracted efficiently, so that changes can also be inventoried and monitored. In this paper, a multisource CD methodology for multiresolution datasets is applied. First, different change indices are processed, then different thresholding algorithms for change/no_change are applied to these indices in order to better estimate the statistical parameters of these categories, finally the indices are integrated into a change detection multisource fusion process, which allows generating a single CD result from several combination of indices. This methodology has been applied to datasets with different spectral and spatial resolution properties. Then, the obtained results are evaluated by means of a quality control analysis, as well as with complementary graphical representations. The suggested methodology has also been proved efficiently for identifying the change detection index with the higher contribution.

Digital image sensor-based assessment of the status of oat (Avena sativa L.) crops after frost damage.

The aim of this paper is to classify the land covered with oat crops, and the quantification of frost damage on oats, while plants are still in the flowering stage. The images are taken by a digital colour camera CCD-based sensor. Unsupervised classification methods are applied because the plants present different spectral signatures, depending on two main factors: illumination and the affected state. The colour space used in this application is CIELab, based on the decomposition of the colour in three channels, because it is the closest to human colour perception. The histogram of each channel is successively split into regions by thresholding. The best threshold to be applied is automatically obtained as a combination of three thresholding strategies: (a) Otsu's method, (b) Isodata algorithm, and (c) Fuzzy thresholding. The fusion of these automatic thresholding techniques and the design of the classification strategy are some of the main findings of the paper, which allows an estimation of the damages and a prediction of the oat production.

An image segmentation based on a genetic algorithm for determining soil coverage by crop residues.

Determination of the soil coverage by crop residues after ploughing is a fundamental element of Conservation Agriculture. This paper presents the application of genetic algorithms employed during the fine tuning of the segmentation process of a digital image with the aim of automatically quantifying the residue coverage. In other words, the objective is to achieve a segmentation that would permit the discrimination of the texture of the residue so that the output of the segmentation process is a binary image in which residue zones are isolated from the rest. The RGB images used come from a sample of images in which sections of terrain were photographed with a conventional camera positioned in zenith orientation atop a tripod. The images were taken outdoors under uncontrolled lighting conditions. Up to 92% similarity was achieved between the images obtained by the segmentation process proposed in this paper and the templates made by an elaborate manual tracing process. In addition to the proposed segmentation procedure and the fine tuning procedure that was developed, a global quantification of the soil coverage by residues for the sampled area was achieved that differed by only 0.85% from the quantification obtained using template images. Moreover, the proposed method does not depend on the type of residue present in the image. The study was conducted at the experimental farm "El Encín" in Alcalá de Henares (Madrid, Spain).

Mapping wide row crops with video sequences acquired from a tractor moving at treatment speed.

This paper presents a mapping method for wide row crop fields. The resulting map shows the crop rows and weeds present in the inter-row spacing. Because field videos are acquired with a camera mounted on top of an agricultural vehicle, a method for image sequence stabilization was needed and consequently designed and developed. The proposed stabilization method uses the centers of some crop rows in the image sequence as features to be tracked, which compensates for the lateral movement (sway) of the camera and leaves the pitch unchanged. A region of interest is selected using the tracked features, and an inverse perspective technique transforms the selected region into a bird's-eye view that is centered on the image and that enables map generation. The algorithm developed has been tested on several video sequences of different fields recorded at different times and under different lighting conditions, with good initial results. Indeed, lateral displacements of up to 66% of the inter-row spacing were suppressed through the stabilization process, and crop rows in the resulting maps appear straight.

A stereovision matching strategy for images captured with fish-eye lenses in forest environments.

We present a novel strategy for computing disparity maps from hemispherical stereo images obtained with fish-eye lenses in forest environments. At a first segmentation stage, the method identifies textures of interest to be either matched or discarded. This is achieved by applying a pattern recognition strategy based on the combination of two classifiers: Fuzzy Clustering and Bayesian. At a second stage, a stereovision matching process is performed based on the application of four stereovision matching constraints: epipolar, similarity, uniqueness and smoothness. The epipolar constraint guides the process. The similarity and uniqueness are mapped through a decision making strategy based on a weighted fuzzy similarity approach, obtaining a disparity map. This map is later filtered through the Hopfield Neural Network framework by considering the smoothness constraint. The combination of the segmentation and stereovision matching approaches makes the main contribution. The method is compared against the usage of simple features and combined similarity matching strategies.

A Hopfield Neural Network for combining classifiers applied to textured images.

In this paper we propose a new method for combining simple classifiers through the analogue Hopfield Neural Network (HNN) optimization paradigm for classifying natural textures in images. The base classifiers are the Fuzzy clustering (FC) and the parametric Bayesian estimator (BP). An initial unsupervised training phase determines the number of clusters and estimates the parameters for both FC and BP. Then a decision phase is carried out, where we build as many Hopfield Neural Networks as the available number of clusters. The number of nodes at each network is the number of pixels in the image which is to be classified. Each node at each network is initially loaded with a state value, which is the membership degree (provided by FC) that the node (pixel) belongs to the cluster associated to the network. Each state is later iteratively updated during the HNN optimization process taking into account the previous states and two types of external influences exerted by other nodes in its neighborhood. The external influences are mapped as consistencies. One is embedded in an energy term which considers the states of the node to be updated and the states of its neighbors. The other is mapped as the inter-connection weights between the nodes. From BP, we obtain the probabilities that the nodes (pixels) belong to a cluster (network). We define these weights as a relation between states and probabilities between the nodes in the neighborhood of the node which is being updated. This is the classifier combination, making the main finding of this paper. The proposed combined strategy based on the HNN outperforms the simple classifiers and also classical combination strategies.