Convolutional Neural Network Regression Research Articles

Deep Learning to Predict Neonatal and Infant Brain Age from Myelination on Brain MRI Scans.

Background Assessment of appropriate brain myelination on T1- and T2-weighted MRI scans is based on gestationally corrected age (GCA) and requires subjective visual inspection of the brain with knowledge of normal myelination milestones. Purpose To develop a convolutional neural network (CNN) capable of estimating neonatal and infant GCA based on brain myelination on MRI scans. Materials and methods In this retrospective study from one academic medical center, brain MRI scans of patients aged 0-25 months with reported normal myelination were consecutively collected between January 1995 and June 2019. The GCA at MRI was manually calculated. After exclusion criteria were applied, T1- and T2-weighted MRI scans were preprocessed with skull stripping, linear registration, z scoring for normalization, and downsampling. A three-dimensional regression CNN was trained to predict GCA using mean absolute error (MAE) as its loss function. Attention maps were calculated using layer-wise relevance propagation. Models were validated on an external test set from the National Institutes of Health (NIH). Model MAEs were compared using Kruskal-Wallis and Mann-Whitney tests. Results A total of 518 neonates and infants (mean GCA, 67 weeks ± 33 [SD], 56% male) was included, comprising 469 T1-, 438 T2-, and 389 T1- and T2-weighted studies. Across 10 runs, MAEs of T1-, T2-, and T1- and T2-weighted networks were 9.8 ± 2.3, 9.1 ± 1.9, and 7.7 ± 1.7 weeks, respectively. Attention map analysis demonstrated increased network attention to the cerebellum, posterior white matter, and basal ganglia signal in neonates with GCA of less than 40 weeks and the anterior white matter signal in infants with GCA of more than 120 weeks, corresponding to the known progression of myelination. The T1- and T2-weighted network tested on the external NIH test set had an MAE of 9.1 weeks, which was reduced to 5.9 weeks with further training using half the external test set (P < .001). Conclusion A three-dimensional convolutional neural network can predict the gestationally corrected age of neonates and infants aged 0-25 months based on brain myelination patterns on T1- and T2-weighted MRI scans. © RSNA, 2022 Online supplemental material is available for this article.

3D Facial Landmark Localization for cephalometric analysis.

Cephalometric analysis is an important and routine task in the medical field to assess craniofacial development and to diagnose cranial deformities and midline facial abnormalities. The advance of 3D digital techniques potentiated the development of 3D cephalometry, which includes the localization of cephalometric landmarks in the 3D models. However, manual labeling is still applied, being a tedious and time-consuming task, highly prone to intra/inter-observer variability. In this paper, a framework to automatically locate cephalometric landmarks in 3D facial models is presented. The landmark detector is divided into two stages: (i) creation of 2D maps representative of the 3D model; and (ii) landmarks' detection through a regression convolutional neural network (CNN). In the first step, the 3D facial model is transformed to 2D maps retrieved from 3D shape descriptors. In the second stage, a CNN is used to estimate a probability map for each landmark using the 2D representations as input. The detection method was evaluated in three different datasets of 3D facial models, namely the Texas 3DFR, the BU3DFE, and the Bosphorus databases. An average distance error of 2.3, 3.0, and 3.2 mm were obtained for the landmarks evaluated on each dataset. The obtained results demonstrated the accuracy of the method in different 3D facial datasets with a performance competitive to the state-of-the-art methods, allowing to prove its versability to different 3D models. Clinical Relevance- Overall, the performance of the landmark detector demonstrated its potential to be used for 3D cephalometric analysis.

Prediction of oil content in single maize kernel based on hyperspectral imaging and attention convolution neural network

An attention (A) based convolutional neural network regression (CNNR) model, namely ACNNR, was proposed to combine hyperspectral imaging to predict oil content in single maize kernel. During the period, a reflectance HSI system was used to collect hyperspectral images of embryo side and non-embryo side of single maize kernel, and the performances of CNNR (without attention mechanism), ACNNR and partial least squares regression (PLSR) were compared. For PLSR, a series of spectral preprocessing and dimensionality reduction methods were used to finally determine the optimal hybrid PLSR model. Whereas for CNNR and ACNNR, only raw spectra were used as their inputs. The results showed that embryo side was more suitable for developing regression models; the attentional mechanism was helpful to reduce the error of prediction, making ACNNR performed best (coefficient of determination of prediction = 0.9198). Overall, the proposed method did not require additional processing on raw spectra, and performed well.

Noise reduction technique using deep learning for ultrasound imaging during high-intensity focused ultrasound treatment

One of the problems with ultrasound imaging during high-intensity focused ultrasound (HIFU) treatment is that the therapeutic ultrasound components interfere with the diagnostic ultrasound components, making it impossible to monitor the tissue changes during HIFU exposure. In this study, a convolutional neural network (CNN) framework was applied to the reconstructed ultrasound images with HIFU noise to remove the therapeutic ultrasound components while the diagnostic ultrasound components remain intact. In the experiments, the chicken breast was used as a tissue sample and exposed to HIFU in the water tank. The ultrasound images with and without noise were acquired during an intermission period of HIFU exposure and the noise-reduced images was predicted using the proposed multi-layer regression CNN model through the training process. As a result, ultrasound images with sufficient spatial resolution to detect the thermal lesion were acquired.

Double Fourier Integral Analysis Based Convolutional Neural Network Regression for High-Frequency Energy Disaggregation

Non-Intrusive Load Monitoring aims to extract the energy consumption of individual electrical appliances through disaggregation of the total power load measured by a single smart-meter. In this article we introduce Double Fourier Integral Analysis in the Non-Intrusive Load Monitoring task in order to provide more distinct feature descriptions compared to current or voltage spectrograms. Specifically, the high-frequency aggregated current and voltage signals are transformed into two-dimensional unit cells as calculated by Double Fourier Integral Analysis and used as input to a Convolutional Neural Network for regression. The performance of the proposed methodology was evaluated in the publicly available U.K.-DALE dataset. The proposed approach improves the estimation accuracy by 7.2&#x0025; when compared to the baseline energy disaggregation setup using current and voltage spectrograms.

A Two-Stage Transfer Regression Convolutional Neural Network for Bearing Remaining Useful Life Prediction

Recently, deep learning techniques have been successfully used for bearing remaining useful life (RUL) prediction. However, the degradation pattern of bearings can be much different from each other, which leads to the trained model usually not being able to work well for RUL prediction of a new bearing. As a method that can adapt a model trained on source datasets to a different but relative unlabeled target dataset, transfer learning shows the potential to solve this problem. Therefore, we propose a two-stage transfer regression (TR)-based bearing RUL prediction method. Firstly, the incipient fault point (IFP) is detected by a convolutional neural network (CNN) classifier to identity the start time of degradation stage and label the training samples. Then, a transfer regression CNN with multiloss is constructed for RUL prediction, including regression loss, classification loss, maximum mean discrepancy (MMD) and regularization loss, which can not only extract fault information from fault classification loss for RUL prediction, but also minimize the probability distribution distance, thus helping the method to be trained in a domain-invariant way via the transfer regression algorithm. Finally, real data collected from run-to-failure bearing experiments are analyzed by the TR-based CNN method. The results and comparisons with state-of-the-art methods demonstrate the superiority and reliable performance of the proposed method for bearing RUL prediction.

Deep learning-based plane pose regression in obstetric ultrasound

PurposeIn obstetric ultrasound (US) scanning, the learner’s ability to mentally build a three-dimensional (3D) map of the fetus from a two-dimensional (2D) US image represents a major challenge in skill acquisition. We aim to build a US plane localisation system for 3D visualisation, training, and guidance without integrating additional sensors.MethodsWe propose a regression convolutional neural network (CNN) using image features to estimate the six-dimensional pose of arbitrarily oriented US planes relative to the fetal brain centre. The network was trained on synthetic images acquired from phantom 3D US volumes and fine-tuned on real scans. Training data was generated by slicing US volumes into imaging planes in Unity at random coordinates and more densely around the standard transventricular (TV) plane.ResultsWith phantom data, the median errors are 0.90 mm/1.17^circ and 0.44 mm/1.21^circ for random planes and planes close to the TV one, respectively. With real data, using a different fetus with the same gestational age (GA), these errors are 11.84 mm/25.17^circ . The average inference time is 2.97 ms per plane.ConclusionThe proposed network reliably localises US planes within the fetal brain in phantom data and successfully generalises pose regression for an unseen fetal brain from a similar GA as in training. Future development will expand the prediction to volumes of the whole fetus and assess its potential for vision-based, freehand US-assisted navigation when acquiring standard fetal planes.

Hyperspectral imaging coupled with CNN: A powerful approach for quantitative identification of feather meal and fish by-product meal adulterated in marine fishmeal

Marine fishmeal (MFM) adulterated with low-cost processed animal proteins (PAPs) such as hydrolyzed feather meal (HFM) and fish by-product meal (FBM) has frequently occurred in the Chinese trade market. This commercial fraud generates a serious threat to farmed animal health and even human food safety. This study aims to develop a rapid detection method using near-infrared hyperspectral imaging (NIR-HSI) combined with deep learning modeling for qualitative and quantitative identification of MFM adulterated with HFM, FBM, and the binary adulterant (HFM-FBM). Three convolutional neural network (CNN) architectures with optimized parameters were constructed to predict sample classes, adulterant concentration, and amino acid content of adulterated samples, respectively. Partial least squares (PLS) and support vector machine (SVM) models were compared with the proposed CNN models. The overall results showed that the CNN outperformed the PLS and SVM on both classification and regression. The six-classification accuracy obtained by the CNN was up to 99.37%, while the R2 of CNN regression prediction varied from 0.984 to 0.997. This study demonstrates that NIR-HSI coupled with CNN calibration provides a promising technique for the detection of MFM adulterated with PAPs.

Automated porosity estimation using CT-scans of extracted core data

Estimation of porosity at a millimeter scale would be an order of magnitude finer resolution than traditional logging techniques. This enables proper description of reservoirs with thin layers and fine scale heterogeneities. To achieve this, we propose an end-to-end convolutional neural network (CNN) regression model that automatically predicts continuous porosity at a millimeter scale resolution using two-dimensional whole core CT scan images. More specifically, a CNN regression model is trained to learn from routine core analysis (RCA) porosity measurements. To characterize the performance of such approach, we compare the performance of this model with two linear regression models trained to learn the relationship between the average attenuation and standard deviation of the same two-dimensional images and RCA porosity. Our investigations reveal that the linear models are outperformed by the CNN, indicating the capability of the CNN model in extracting textures that are important for porosity estimations. We compare the predicted porosity results against the total porosity logs calculated from the density log. The obtained results show that the predicted porosity values using the proposed CNN method are well correlated with the core plug measurements and the porosity log. More importantly, the proposed approach can provide accurate millimeter scale porosity estimations, while the total porosity log is averaged over an interval and thus do not show such fine scale variations. Thus, the proposed method can be employed to calibrate the porosity logs, thereby reducing the uncertainties associated with indirect calculations of the porosity from such logs.

A fuzzy convolutional neural network for enhancing multi-focus image fusion

The images captured by the cameras contain distortions, misclassified pixels, uncertainties and poor contrast. Therefore, the multi-focus image fusion (MFIF) integrates various input image features to produce a single fused image using all its objects in focus. However, it is computationally complex, which leads to inconsistency. Hence, the MFIF method is employed to generate the fused image by integrating the fuzzy sets (FS) and convolutional neural network (CNN) to detect focused and unfocused parts in both source images. It is also compared with other competing six MFIF methods like Neutrosophic set based stationary wavelet transform (NSWT), guided filters, CNN, ensemble CNN, image fusion-based CNN and deep regression pair learning (DRPL). Benchmark datasets validate the superiority of the proposed FCNN method in terms of four non-reference assessment measures having mutual information (1.1678), edge information (0.7281), structural similarity (0.9850) and human perception (0.8020) and two reference metrics such as Peak signal-to-noise ratio (57.23) and root mean square error (1.814).

Quantitative estimation of pulmonary artery wedge pressure from chest radiographs by a regression convolutional neural network

Recent studies reported that a convolutional neural network (CNN; a deep learning model) can detect elevated pulmonary artery wedge pressure (PAWP) from chest radiographs, the diagnostic images most commonly used for assessing pulmonary congestion in heart failure. However, no method has been published for quantitatively estimating PAWP from such radiographs. We hypothesized that a regression CNN, an alternative type of deep learning, could be a useful tool for quantitatively estimating PAWP in cardiovascular diseases. We retrospectively enrolled 936 patients with cardiovascular diseases who had undergone right heart catheterization (RHC) and chest radiography and estimated PAWP by constructing a regression CNN based on the VGG16 model. We randomly categorized 80% of the data as training data (training group, n = 748) and 20% as test data (test group, n = 188). Moreover, we tuned the learning rate—one of the model parameters—by 5-hold cross-validation of the training group. Correlations between PAWP measured by RHC [ground truth (GT) PAWP] and PAWP derived from the regression CNN (estimated PAWP) were tested. To visualize how the regression CNN assessed the images, we created a regression activation map (RAM), a visualization technique for regression CNN. Estimated PAWP correlated significantly with GT PAWP in both the training (r = 0.76, P < 0.001) and test group (r = 0.62, P < 0.001). Bland–Altman plots found a mean (SEM) difference between GT and estimated PAWP of − 0.23 (0.16) mm Hg in the training and − 0.05 (0.41) mm Hg in the test group. The RAM showed that our regression CNN model estimated high PAWP by focusing on the cardiomegaly and pulmonary congestion. In the test group, the area under the curve (AUC) for detecting elevated PAWP (≥ 18 mm Hg) produced by the regression CNN model was similar to the AUC of an experienced cardiologist (0.86 vs 0.83, respectively; P = 0.24). This proof-of-concept study shows that regression CNN can quantitatively estimate PAWP from standard chest radiographs in cardiovascular diseases.

Accurate and rapid prediction of tuberculosis drug resistance from genome sequence data using traditional machine learning algorithms and CNN

Effective and timely antibiotic treatment depends on accurate and rapid in silico antimicrobial-resistant (AMR) predictions. Existing statistical rule-based Mycobacterium tuberculosis (MTB) drug resistance prediction methods using bacterial genomic sequencing data often achieve varying results: high accuracy on some antibiotics but relatively low accuracy on others. Traditional machine learning (ML) approaches have been applied to classify drug resistance for MTB and have shown more stable performance. However, there is no study that uses deep learning architecture like Convolutional Neural Network (CNN) on a large and diverse cohort of MTB samples for AMR prediction. We developed 24 binary classifiers of MTB drug resistance status across eight anti-MTB drugs and three different ML algorithms: logistic regression, random forest and 1D CNN using a training dataset of 10,575 MTB isolates collected from 16 countries across six continents, where an extended pan-genome reference was used for detecting genetic features. Our 1D CNN architecture was designed to integrate both sequential and non-sequential features. In terms of F1-scores, 1D CNN models are our best classifiers that are also more accurate and stable than the state-of-the-art rule-based tool Mykrobe predictor (81.1 to 93.8%, 93.7 to 96.2%, 93.1 to 94.8%, 95.9 to 97.2% and 97.1 to 98.2% for ethambutol, rifampicin, pyrazinamide, isoniazid and ofloxacin respectively). We applied filter-based feature selection to find AMR relevant features. All selected variant features are AMR-related ones in CARD database. 78.8% of them are also in the catalogue of MTB mutations that were recently identified as drug resistance-associated ones by WHO. To facilitate ML model development for AMR prediction, we packaged every step into an automated pipeline and shared the source code at https://github.com/KuangXY3/MTB-AMR-classification-CNN.

An improved AIC onset-time picking method based on regression convolutional neural network

Akaike information criterion, known as AIC, has become one of the most used methods for acoustic emission (AE) signals onset-time picking since it was proposed in 1970s. However in practical applications, the automatic onset-times picking are hard to perform precisely due to the interference of the strong background noise and static noise, which affects the accuracy of AIC picking. In this work, an improved AIC onset-time picking method based on regression convolutional neural network (CNN) is proposed. First, several features of AE signals to be trained are selected manually, and arrival times of AE signals are labeled correspondingly. Then datasets with features and labels are put into the regression CNN model for training and enhancing the connection of the signals in the time domain. Finally, AIC algorithm is applied to obtain the onset times of the signals processed by the trained CNN model. Based on the Hsu-Nielsen source AE data, the stability and performance of the proposed method are tested, analyzed and compared with those of other mainstream detection methods: AIC, short/long term average combined with AIC (STA/LTA-AIC), and floating threshold (FT). The results prove that the accuracy of the proposed method significantly exceeds that of other methods. Meanwhile, especially in low signal-to-noise ratios (SNRs) scenario, the accuracy stability of the improved method has excellent accuracy and stability, which proves that the proposed method has promising onset-time picking performance for AE signals, including signals with low SNR characteristics.

Position and Pose Measurement of 3-PRS Ankle Rehabilitation Robot Based on Deep Learning

Introduction: The position and pose measurement of the rehabilitation robot plays a very important role in patient rehabilitation movement, and the non-contact real-time robot position and pose measurement is of great significance. Rehabilitation training is a relatively complicated process, so it is very important to detect the training process of the rehabilitation robot in real time and accuracy. The method of the deep learning has a very good effect on monitoring the rehabilitation robot state. Methods: The structure sketch and the 3D model of the 3-PRS ankle rehabilitation robot are established, and the mechanism kinematics is analyzed to obtain the relationship between the driving input - the three slider heights - and the position and pose parameters. The whole network of the position and pose measurement is composed of two stages: (1) measuring the slider heights using the CNN based on the robot image and (2) calculating the position and pose parameter using the BPNN based on the measured slider heights from the CNN. According to the characteristics of continuous variation of the slider heights, a regression CNN is proposed and established to measure the robot slider height. Based on the data calculated by using the inverse kinematics of the 3-PRS ankle rehabilitation robot, a BPNN is established to solve the forward kinematics for the position and pose. Results: The experimental results show that the regression CNN outputs the slider height and then the BPNN accurately outputs the corresponding position and pose. Eventually, the position and pose parameters are obtained from the robot image. Compared with the traditional robot position and pose measurement method, the proposed method has significant advantages. Conclusion: The proposed 3-PRS ankle rehabilitation position and pose method can not only shorten the experiment period and cost, but also get excellent timeliness and precision. The proposed approach can help the medical staff to monitor the status of the rehabilitation robot and help the patient rehabilitation in training. Discussion: The goal of the work is to construct a new position and pose detection network based on the combination of the regression convolutional neural network (CNN) and the back propagation neural network (BPNN). The main contribution is to measure the position and pose of the 3-PRS ankle rehabilitation robot in real time, which improves the measurement accuracy and the efficiency of the medical staff work.

Segmentation-Based vs. Regression-Based Biomarker Estimation: A Case Study of Fetus Head Circumference Assessment from Ultrasound Images.

The fetus head circumference (HC) is a key biometric to monitor fetus growth during pregnancy, which is estimated from ultrasound (US) images. The standard approach to automatically measure the HC is to use a segmentation network to segment the skull, and then estimate the head contour length from the segmentation map via ellipse fitting, usually after post-processing. In this application, segmentation is just an intermediate step to the estimation of a parameter of interest. Another possibility is to estimate directly the HC with a regression network. Even if this type of segmentation-free approaches have been boosted with deep learning, it is not yet clear how well direct approach can compare to segmentation approaches, which are expected to be still more accurate. This observation motivates the present study, where we propose a fair, quantitative comparison of segmentation-based and segmentation-free (i.e., regression) approaches to estimate how far regression-based approaches stand from segmentation approaches. We experiment various convolutional neural networks (CNN) architectures and backbones for both segmentation and regression models and provide estimation results on the HC18 dataset, as well agreement analysis, to support our findings. We also investigate memory usage and computational efficiency to compare both types of approaches. The experimental results demonstrate that even if segmentation-based approaches deliver the most accurate results, regression CNN approaches are actually learning to find prominent features, leading to promising yet improvable HC estimation results.

The value of probabilistic forecasting in emergency medical resource planning under uncertainty

PurposeMost epidemic transmission forecasting methods can only provide deterministic outputs. This study aims to show that probabilistic forecasting, in contrast, is suitable for stochastic demand modeling and emergency medical resource planning under uncertainty.Design/methodology/approachTwo probabilistic forecasting methods, i.e. quantile regression convolutional neural network and kernel density estimation, are combined to provide the conditional quantiles and conditional densities of infected populations. The value of probabilistic forecasting in improving decision performances and controlling decision risks is investigated by an empirical study on the emergency medical resource planning for the COVID-19 pandemic.FindingsThe managerial implications obtained from the empirical results include (1) the optimization models using the conditional quantile or the point forecasting result obtain better results than those using the conditional density; (2) for sufficient resources, decision-makers' risk preferences can be incorporated to make tradeoffs between the possible surpluses and shortages of resources in the emergency medical resource planning at different quantile levels; and (3) for scarce resources, the differences in emergency medical resource planning at different quantile levels greatly decrease or disappear because of the existing of forecasting errors and supply quantity constraints.Originality/valueVery few studies concern probabilistic epidemic transmission forecasting methods, and this is the first attempt to incorporate deep learning methods into a two-phase framework for data-driven emergency medical resource planning under uncertainty. Moreover, the findings from the empirical results are valuable to select a suitable forecasting method and design an efficient emergency medical resource plan.

SSRNet: In-Field Counting Wheat Ears Using Multi-Stage Convolutional Neural Network

Fast and accurate counting of wheat ears in field conditions is a key element for determining wheat yield. To obtain the number of wheat ears in a field, we propose a new counting algorithm based on computer vision. This algorithm counts wheat ears in remote images through semantic segmentation regression network (SSRNet). SSRNet is a multistage convolutional neural network that we propose to achieve counting problems through regression. In SSRNet, first, the original image is cropped to increase the amount of data. This method effectively solves the small sample dataset. Next, based on the cropping results, we build a fully convolutional neural network (FCNN) to segment wheat ears in field conditions. FCNN increases the accuracy of wheat ears counting by accurately segmenting wheat ears in a complex background. Then, we build a regression convolutional neural network (RCNN) to count wheat ears based on the segmentation results of FCNN. In RCNN, we propose a new activation function positive rectification linear unit (PrLU) to process the last layer of the fully connected layer, so that RCNN can effectively count the number of wheat ears in the image. Finally, a counting strategy is proposed to count the number of wheat ears in the original image. To verify the counting performance of SSRNet, we compare the counting result of SSRNet with the real value of manual statistics. The results show that the average accuracy (Acc), <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula>, and root mean squared error (RMSE) of the SSRNet count results on the test set in this article are 0.980, 0.996, and 9.437, respectively. It can be seen from the results that our proposed method can accurately count wheat ears in field conditions. At the same time, the counting time (0.11 s) shows that SSRNet can quickly estimate the number of wheat ears in field conditions. We concluded that this study can provide important technical support for the high-throughput field wheat ears counting task in large-scale phenotyping work.

Orientation Estimation of Rotated Sonar Image Targets via the Wavelet Subimage Energy Ratio

Precise orientation information facilitates target recognition of sonar images. Traditional orientation estimation methods, including the Hough transform method, have poor antinoise abilities, so they cannot achieve high estimation precision for sonar images, which usually have low signal-to-noise ratios. The convolutional neural network (CNN) regression method is sensitive to image affine transformations and requires extensive computation, so it cannot achieve strong robustness or high speed. To achieve high-precision, high-speed, and robust orientation estimation of sonar image targets, we present a novel orientation estimation method via the wavelet subimage energy ratio (WSER). The WSER varies with the rotation angles of images and has the highest value when the long axes of targets are vertical. It is translational and scale invariant and does not need supervised training. Therefore, we propose estimating orientations of sonar image targets by finding the max of the WSER to achieve high precision, high speed, and strong robustness to the translation and scale transformations of images. The results of experiments on a self-made sonar image dataset show that the mean absolute error (MAE) of the proposed method is 7.9°, while the MAEs of the CNN regression method and traditional methods are 10.1 and 22.9°, respectively. In addition, the proposed method is six times faster than the CNN regression method. The proposed orientation estimation method has also been applied to align the orientation of sonar images and CNN can achieve state-of-the-art performance for rotated target recognition using the aligned images. Data and codes are publicly available.

A CNN-Regression-Based Contact Erosion Measurement Method for AC Contactors

This paper proposes a method that indirectly measures the contact erosion of alternating current (AC) contactors via mapping electrical signals to the contacting alloy erosion condition which is represented by the accumulated contact mass loss (ACML). Electrical signal waveforms and their corresponding ACMLs of every make-and-break operation are acquired in endurance tests. A supervised convolutional neural network regression (CNNR) architecture containing six one-dimensional convolution layers is proposed to model the relation between waveforms and ACMLs. We compare different CNNR architectures as well as different training schemes by the test precision to obtain the optimal solution. Experiments prove that the proposed CNNR architecture with an optimized training scheme can achieve a precise ACML measurement when only voltage waveforms of make operations are used. The best results reach mean absolute errors of 3.29% and 1.59% corresponding to two datasets respectively, which are superior to the results of other regression methods in the comparison and prove the theoretical significance and applicational values.

Deep convolutional neural network for damaged vegetation segmentation from RGB images based on virtual NIR-channel estimation

Performing accurate and automated semantic segmentation of vegetation is a first algorithmic step towards more complex models that can extract accurate biological information on crop health, weed presence and phenological state, among others. Traditionally, models based on normalized difference vegetation index (NDVI), near infrared channel (NIR) or RGB have been a good indicator of vegetation presence. However, these methods are not suitable for accurately segmenting vegetation showing damage, which precludes their use for downstream phenotyping algorithms. In this paper, we propose a comprehensive method for robust vegetation segmentation in RGB images that can cope with damaged vegetation. The method consists of a first regression convolutional neural network to estimate a virtual NIR channel from an RGB image. Second, we compute two newly proposed vegetation indices from this estimated virtual NIR: the infrared-dark channel subtraction (IDCS) and infrared-dark channel ratio (IDCR) indices. Finally, both the RGB image and the estimated indices are fed into a semantic segmentation deep convolutional neural network to train a model to segment vegetation regardless of damage or condition. The model was tested on 84 plots containing thirteen vegetation species showing different degrees of damage and acquired over 28 days. The results show that the best segmentation is obtained when the input image is augmented with the proposed virtual NIR channel (F1=0.94) and with the proposed IDCR and IDCS vegetation indices (F1=0.95) derived from the estimated NIR channel, while the use of only the image or RGB indices lead to inferior performance (RGB(F1=0.90) NIR(F1=0.82) or NDVI(F1=0.89) channel). The proposed method provides an end-to-end land cover map segmentation method directly from simple RGB images and has been successfully validated in real field conditions.