Automatic strawberry leaf scorch severity estimation via faster R-CNN and few-shot learning

Using deep learning to estimate strawberry leaf scorch severity often achieves unsatisfactory results when a strawberry leaf image contains complex background information or multi-class diseased leaves and the number of annotated strawberry leaf images is limited. To solve these issues, in this paper, we propose a two-stage method including object detection and few-shot learning to estimate strawberry leaf scorch severity. In the first stage, Faster R-CNN is used to mark the location of strawberry leaf patches, where each single strawberry leaf patch is clipped from original strawberry leaf images to compose a new strawberry leaf patch dataset. In the second stage, the Siamese network trained on the new strawberry leaf patch dataset is used to identify the strawberry leaf patches and then estimate the severity of the original strawberry leaf scorch images according to the multi-instance learning concept. Experimental results from the first stage show that Faster R-CNN achieves better mAP in strawberry leaf patch detection than other object detection networks, at 94.56%. Results from the second stage reveal that the Siamese network achieves an accuracy of 96.67% in the identification of strawberry disease leaf patches, which is higher than the Prototype network. Comprehensive experimental results indicate that compared with other state-of-the-art models, our proposed two-stage method comprising the Faster R-CNN (VGG16) and Siamese networks achieves the highest estimation accuracy of 96.67%. Moreover, our trained two-stage model achieves an estimation accuracy of 88.83% on a new dataset containing 60 strawberry leaf images taken in the field, which indicates its excellent generalization ability.     

Semantic region of interest and species classification in the deep neural network feature domain

In this paper, we focus on animal object detection and species classification in camera-trap images collected in highly cluttered natural scenes. Using a deep neural network (DNN) model training for animal- background image classification, we analyze the input camera-trap images to generate a multi-level visual representation of the input image. We detect semantic regions of interest for animals from this representation using k-mean clustering and graph cut in the DNN feature domain. These animal regions are then classified into animal species using multi-class deep neural network model. According the experimental results, our method achieves 99.75% accuracy for classifying animals and background and 90.89% accuracy for classifying 26 animal species on the Snapshot Serengeti dataset, outperforming existing image classification methods.     

Real-time fish detection in complex backgrounds using probabilistic background modelling

Computer vision and image processing approaches for automatic underwater fish detection are gaining attention of marine scientists as quicker and low-cost methods for estimating fish biomass and assemblage in oceans and fresh water bodies. However, the main challenge that is encountered in unconstrained underwater imagery is poor luminosity, turbidity, background confusion and foreground camouflage that make conventional approaches compromise on their performance due to missed detections or high false alarm rates. Gaussian Mixture Modelling is a powerful approach to segment foreground fish from the background objects through learning the background pixel distribution. In this paper, we present an algorithm based on Gaussian Mixture Models together with Pixel-Wise Posteriors for fish detection in complex background scenarios. We report the results of our method on the benchmark Complex Background dataset that is extracted from Fish4Knowledge repository. Our proposed method yields an F-score of 84.3%, which is the highest score reported so far on the aforementioned dataset for detecting fish in an unconstrained environment.     

基于F3Net显著性目标检测的蝴蝶图像前背景自动分割

[目的]具有复杂背景的蝴蝶图像前背景分割难度大.本研究旨在探索基于深度学习显著性目标检测的蝴蝶图像自动分割方法.[方法]应用DUTS-TR数据集训练F3Net显著性目标检测算法构建前背景预测模型,然后将模型用于具有复杂背景的蝴蝶图像数据集实现蝴蝶前背景自动分割.在此基础上,采用迁移学习方法,保持ResNet骨架不变,利...     

Capacitance detection in the wave-type electric fish Eigenmannia during active electrolocation

Weakly electric fish can detect nearby objects and analyse their electric properties during active electrolocation. Four individuals of the South American gymnotiform fish Eigenmannia sp., which emits a continuous wave-type electric signal, were tested for their ability to detect capacitive properties of objects and discriminate them from resistive properties. For individual fish, capacitive values of objects had to be greater than 0.22–1.7 nF (`lower threshold') and smaller than 120–680 nF (`upper threshold') in order to be detected. The capacitive values of natural objects fall well within this detection range. All fish trained could discriminate unequivocally between capacitive and resistive object properties. Thus, fish perceive capacitive properties as a separate object quality. The effects of different types of objects on the locally occurring electric signals which stimulate electroreceptors during electrolocation were examined. Purely resistive objects altered mainly local electric organ discharge (EOD) amplitude, but capacitive objects with values between about 0.5 and 600 nF changed the timing of certain EOD parameters (phase-shift) and EOD waveform. A mechanism for capacitance detection in wave-type electric fish based on time measurements is proposed and compared with the capacitance detection mechanism in mormyrid pulse-type fish, which is based on waveform measurements. Accepted: 31 July 1997     

Using a binaural biomimetic array to identify bottom objects ensonified by echolocating dolphins

The development of a unique dolphin biomimetic sonar produced data that were used to study signal processing methods for object identification. Echoes from four metallic objects proud on the bottom, and a substrate-only condition, were generated by bottlenose dolphins trained to ensonify the targets in very shallow water. Using the two-element ('binaural') receive array, object echo spectra were collected and submitted for identification to four neural network architectures. Identification accuracy was evaluated over two receive array configurations, and five signal processing schemes. The four neural networks included backpropagation, learning vector quantization, genetic learning and probabilistic network architectures. The processing schemes included four methods that capitalized on the binaural data, plus a monaural benchmark process. All the schemes resulted in above-chance identification accuracy when applied to learning vector quantization and backpropagation. Beam-forming or concatenation of spectra from both receive elements outperformed the monaural benchmark, with higher sensitivity and lower bias. Ultimately, best object identification performance was achieved by the learning vector quantization network supplied with beam-formed data. The advantages of multi-element signal processing for object identification are clearly demonstrated in this development of a first-ever dolphin biomimetic sonar.     

Animal Scanner: Software for classifying humans,animals, and empty frames in camera trap images

Camera traps are a popular tool to sample animal populations because they are noninvasive, detect a variety of species, and can record many thousands of animal detections per deployment. Cameras are typically set to take bursts of multiple photographs for each detection and are deployed in arrays of dozens or hundreds of sites, often resulting in millions of photographs per study. The task of converting photographs to animal detection records from such large image collections is daunting, and made worse by situations that generate copious empty pictures from false triggers (e.g., camera malfunction or moving vegetation) or pictures of humans. We developed computer vision algorithms to detect and classify moving objects to aid the first step of camera trap image filtering—separating the animal detections from the empty frames and pictures of humans. Our new work couples foreground object segmentation through background subtraction with deep learning classification to provide a fast and accurate scheme for human–animal detection. We provide these programs as both Matlab GUI and command prompt developed with C++. The software reads folders of camera trap images and outputs images annotated with bounding boxes around moving objects and a text file summary of results. This software maintains high accuracy while reducing the execution time by 14 times. It takes about 6 seconds to process a sequence of ten frames (on a 2.6 GHZ CPU computer). For those cameras with excessive empty frames due to camera malfunction or blowing vegetation automatically removes 54% of the false‐triggers sequences without influencing the human/animal sequences. We achieve 99.58% on image‐level empty versus object classification of Serengeti dataset. We offer the first computer vision tool for processing camera trap images providing substantial time savings for processing large image datasets, thus improving our ability to monitor wildlife across large scales with camera traps.     

Background sampling for multi-scale ensemble habitat selection modeling: Does the number of points matter?

Ensemble habitat selection modeling is becoming a popular approach among ecologists to answer different questions. Since we are still in the early stages of development and application of ensemble modeling, there remain many questions regarding performance and parameterization. One important gap, which this paper addresses, is how the number of background points used to train models influences the performance of the ensemble model. We used an empirical presence-only dataset and three different selections of background points to train scale-optimized habitat selection models using six modeling algorithms (GLM, GAM, MARS, ANN, Random Forest, and MaxEnt). We tested four ensemble models using different combinations of the component models: (a) equal numbers of background points and presences, (b) background points equaled ten times the number of presences, (c) 10,000 background points, and (d) optimized background points for each component model. Among regression-based approaches, MARS performed best when built with 10,000 background points. Among machine learning models, RF performed the best when built with equal presences and background points. Among the four ensemble models, AUC indicated that the best performing model was the ensemble with each component model including the optimized number of background points, while TSS increased as the number of background points models increased. We found that an ensemble of models, each trained with an optimal number of background points, outperformed ensembles of models trained with the same number of background points, although differences in performance were slight. When using a single modeling method, RF with equal number of presences and background points can perform better than an ensemble model, but the performance fluctuates when the number of background points is not properly selected. On the other hand, ensemble modeling provides consistently high accuracy regardless of background point sampling approach. Further, optimizing the number of background points for each component model within an ensemble model can provide the best model improvement. We suggest evaluating more models across multiple species to investigate how background point selection might affect ensemble models in different scenarios.     

Fish image recognition method based on multi-layer feature fusion convolutional network

Solving the problem of fish image classification is important to conserve fish diversity. This conundrum can be addressed by developing a new fish image classification method based on deep learning by training data with complex backgrounds. To this end, this paper proposes a fusion model, referred to as Tripmix-Net. The backbone network of the proposed model primarily consists of multiscale parallel and improved residual networks that are connected in an alternate manner, and network fusion is used to integrate the information that is extracted from shallow and deep layers. Experiments conducted on the 15-category WildFish fish image dataset verified the efficacy of the proposed Tripmix-Net for classifying same-genus fish images with complex backgrounds. The model achieved an accuracy of 95.31%, which is a significant improvement over traditional methods. The proposed approach serves as a new concept for the fine-grained image classification of fish against complex backgrounds.     

A novel multistage deep belief network based extreme learning machine ensemble learning paradigm for credit risk assessment

To achieve high assessment accuracy for credit risk, a novel multistage deep belief network (DBN) based extreme learning machine (ELM) ensemble learning methodology is proposed. In the proposed methodology, three main stages, i.e., training subsets generation, individual classifiers training and final ensemble output, are involved. In the first stage, bagging sampling algorithm is applied to generate different training subsets for guaranteeing enough training data. Second, the ELM, an effective AI forecasting tool with the unique merits of time-saving and high accuracy, is utilized as the individual classifier, and diverse ensemble members can be accordingly formulated with different subsets and different initial conditions. In the final stage, the individual results are fused into final classification output via the DBN model with sufficient hidden layers, which can effectively capture the valuable information hidden in ensemble members. For illustration and verification, the experimental study on one publicly available credit risk dataset is conducted, and the results show the superiority of the proposed multistage DBN-based ELM ensemble learning paradigm in terms of high classification accuracy.     

Online learning of objects in a biologically motivated visual architecture

We present a biologically motivated architecture for object recognition that is capable of online learning of several objects based on interaction with a human teacher. The system combines biological principles such as appearance-based representation in topographical feature detection hierarchies and context-driven transfer between different levels of object memory. Training can be performed in an unconstrained environment by presenting objects in front of a stereo camera system and labeling them by speech input. The learning is fully online and thus avoids an artificial separation of the interaction into training and test phases. We demonstrate the performance on a challenging ensemble of 50 objects.     

基于性诱和深度学习的草地贪夜蛾成虫自动识别计数方法

【目的】探究深度学习在草地贪夜蛾Spodoptera frugiperda成虫自动识别计数上的可行性,并评估模型的识别计数准确率,为害虫机器智能监测提供图像识别与计数方法。【方法】设计一种基于性诱的害虫图像监测装置,定时自动采集诱捕到的草地贪夜蛾成虫图像,结合采集船形诱捕器粘虫板上草地贪夜蛾成虫图像,构建数据集;应用YOLOv5深度学习目标检测模型进行特征学习,通过草地贪夜蛾原始图像、清除边缘残缺目标、增加相似检测目标(斜纹夜蛾成虫)、无检测目标负样本等不同处理的数据集进行模型训练,得到Yolov5s-A1, Yolov5s-A2, Yolov5s-AB, Yolov5s-ABC 4个模型,对比在不同遮挡程度梯度下的测试样本不同模型检测结果,用准确率(P)、召回率(R)、F1值、平均准确率(average precision, AP)和计数准确率(counting accuracy, CA)评估各模型的差异。【结果】通过原始图像集训练的模型Yolov5s-A1的识别准确率为87.37%,召回率为90.24%,F1值为88.78;清除边缘残缺目标图像集训练得到的模型Yolov5s-A2的识别准确率为93.15%,召回率为84.77%,F1值为88.76;增加斜纹夜蛾成虫样本图像训练的模型Yolov5s-AB的识别准确率为96.23%,召回率为91.85%,F1值为93.99;增加斜纹夜蛾成虫和无检测对象负样本训练的模型Yolov5s-ABC的识别准确率为94.76%,召回率为88.23%,F1值为91.38。4个模型的AP值从高到低排列如下：Yolov5s-AB>Yolov5s-ABC> Yolov5s-A2>Yolov5s-A1,其中Yolov5s-AB与Yolov5s-ABC结果相近;CA值从高到低排列如下：Yolov5s-AB>Yolov5s-ABC>Yolov5s-A2>Yolov5s-A1。【结论】结果表明本文提出的方法应用于控制条件下害虫图像监测设备及诱捕器粘虫板上草地贪夜蛾成虫的识别计数是可行的,深度学习技术对于草地贪夜蛾成虫的识别和计数是有效的。基于深度学习的草地贪夜蛾成虫自动识别与计数方法对虫体姿态变化、杂物干扰等有较好的鲁棒性,可从各种虫体姿态及破损虫体中自动统计出草地贪夜蛾成虫的数量,在害虫种群监测中具有广阔的应用前景。     

A novel deep learning method for maize disease identification based on small sample-size and complex background datasets

Maize diseases are a major source of yield loss, but due to the lack of human experience and limitations of traditional image-recognition technology, obtaining satisfactory large-scale identification results of maize diseases are difficult. Fortunately, the advancement of deep learning-based technology makes it possible to automatically identify diseases. However, it still faces issues caused by small sample sizes and complex field background, which affect the accuracy of disease identification. To address these issues, a deep learning-based method was proposed for maize disease identification in this paper. DenseNet121 was used as the main extraction network and a multi-dilated-CBAM-DenseNet (MDCDenseNet) model was built by combining the multi-dilated module and convolutional block attention module (CBAM) attention mechanism. Five models of MDCDenseNet, DenseNet121, ResNet50, MobileNetV2, and NASNetMobile were compared and tested using three kinds of maize leave images from the PlantVillage dataset and field-collected at Northeast Agricultural University in China. Furthermore, auxiliary classifier generative adversarial network (ACGAN) and transfer learning were used to expand the dataset and pre-train for optimal identification results. When tested on field-collected datasets with a complex background, the MDCDenseNet model outperformed compared to these models with an accuracy of 98.84%. Therefore, it can provide a viable reference for the identification of maize leaf diseases collected from the farmland with a small sample size and complex background.     

Object detection in the fly during simulated translatory flight

Translatory movement of an animal in its environment induces optic flow that contains information about the three-dimensional layout of the surroundings: as a rule, images of objects that are closer to the animal move faster across the retina than those of more distant objects. Such relative motion cues are used by flies to detect objects in front of a structured background. We confronted flying flies, tethered to a torque meter, with front-to-back motion of patterns displayed on two CRT screens, thereby simulating translatory motion of the background as experienced by an animal during straight flight. The torque meter measured the instantaneous turning responses of the fly around its vertical body axis. During short time intervals, object motion was superimposed on background pattern motion. The average turning response towards such an object depends on both object and background velocity in a characteristic way: (1) in order to elicit significant responses object motion has to be faster than background motion; (2) background motion within a certain range of velocities improves object detection. These properties can be interpreted as adaptations to situations as they occur in natural free flight. We confirmed that the measured responses were mediated mainly by a control system specialized for the detection of objects rather than by the compensatory optomotor system responsible for course stabilization. Accepted: 20 March 1997     

Distance discrimination during active electrolocation in the weakly electric fish Gnathonemus petersii.

Weakly electric fish use active electrolocation for orientation at night. They emit electric signals (electric organ discharges) which generate an electrical field around their body. By sensing field distortions, fish can detect objects and analyze their properties. It is unclear, however, how accurately they can determine the distance of unknown objects. Four Gnathonemus petersii were trained in two-alternative forced-choice procedures to discriminate between two objects differing in their distances to a gate. The fish learned to pass through the gate behind which the corresponding object was farther away. Distance discrimination thresholds for different types of objects were determined. Locomotor and electromotor activity during distance measurement were monitored. Our results revealed that all individuals quickly learned to measure object distance irrespective of size, shape or electrical conductivity of the object material. However, the distances of hollow, water-filled cubes and spheres were consistently misjudged in comparison with solid or more angular objects, being perceived as farther away than they really were. As training continued, fish learned to compensate for these 'electrosensory illusions' and erroneous choices disappeared with time. Distance discrimination thresholds depended on object size and overall object distance. During distance measurement, the fish produced a fast regular rhythm of EOD discharges. A mechanisms for distance determination during active electrolocation is proposed.     

Distance discrimination during active electrolocation in the weakly electric fish Gnathonemus petersii

Weakly electric fish use active electrolocation for orientation at night. They emit electric signals (electric organ discharges) which generate an electrical field around their body. By sensing field distortions, fish can detect objects and analyze their properties. It is unclear, however, how accurately they can determine the distance of unknown objects. Four Gnathonemus petersii were trained in two-alternative forced-choice procedures to discriminate between two objects differing in their distances to a gate. The fish learned to pass through the gate behind which the corresponding object was farther away. Distance discrimination thresholds for different types of objects were determined. Locomotor and electromotor activity during distance measurement were monitored. Our results revealed that all individuals quickly learned to measure object distance irrespective of size, shape or electrical conductivity of the object material. However, the distances of hollow, water-filled cubes and spheres were consistently misjudged in comparison with solid or more angular objects, being perceived as farther away than they really were. As training continued, fish learned to compensate for these 'electrosensory illusions' and erroneous choices disappeared with time. Distance discrimination thresholds depended on object size and overall object distance. During distance measurement, the fish produced a fast regular rhythm of EOD discharges. A mechanisms for distance determination during active electrolocation is proposed.     

A deep learning classifier for digital breast tomosynthesis

PurposeTo develop a computerized detection system for the automatic classification of the presence/absence of mass lesions in digital breast tomosynthesis (DBT) annotated exams, based on a deep convolutional neural network (DCNN).Materials and MethodsThree DCNN architectures working at image-level (DBT slice) were compared: two state-of-the-art pre-trained DCNN architectures (AlexNet and VGG19) customized through transfer learning, and one developed from scratch (DBT-DCNN). To evaluate these DCNN-based architectures we analysed their classification performance on two different datasets provided by two hospital radiology departments. DBT slice images were processed following normalization, background correction and data augmentation procedures. The accuracy, sensitivity, and area-under-the-curve (AUC) values were evaluated on both datasets, using receiver operating characteristic curves. A Grad-CAM technique was also implemented providing an indication of the lesion position in the DBT slice.Results Accuracy, sensitivity and AUC for the investigated DCNN are in-line with the best performance reported in the field. The DBT-DCNN network developed in this work showed an accuracy and a sensitivity of (90% ± 4%) and (96% ± 3%), respectively, with an AUC as good as 0.89 ± 0.04. A k-fold cross validation test (with k = 4) showed an accuracy of 94.0% ± 0.2%, and a F1-score test provided a value as good as 0.93 ± 0.03. Grad-CAM maps show high activation in correspondence of pixels within the tumour regions.Conclusions We developed a deep learning-based framework (DBT-DCNN) to classify DBT images from clinical exams. We investigated also a possible application of the Grad-CAM technique to identify the lesion position.     

Arrhythmic Heartbeat Classification Using 2D Convolutional Neural Networks

BackgroundElectrocardiogram (ECG) is a method of recording the electrical activity of the heart and it provides a diagnostic means for heart-related diseases. Arrhythmia is any irregularity of the heartbeat that causes an abnormality in the heart rhythm. Early detection of arrhythmia has great importance to prevent many diseases. Manual analysis of ECG recordings is not practical for quickly identifying arrhythmias that may cause sudden deaths. Hence, many studies have been presented to develop computer-aided-diagnosis (CAD) systems to automatically identify arrhythmias.MethodsThis paper proposes a novel deep learning approach to identify arrhythmias in ECG signals. The proposed approach identifies arrhythmia classes using Convolutional Neural Network (CNN) trained by two-dimensional (2D) ECG beat images. Firstly, ECG signals, which consist of 5 different arrhythmias, are segmented into heartbeats which are transformed into 2D grayscale images. Afterward, the images are used as input for training a new CNN architecture to classify heartbeats.ResultsThe experimental results show that the classification performance of the proposed approach reaches an overall accuracy of 99.7%, sensitivity of 99.7%, and specificity of 99.22% in the classification of five different ECG arrhythmias. Further, the proposed CNN architecture is compared to other popular CNN architectures such as LeNet and ResNet-50 to evaluate the performance of the study.ConclusionsTest results demonstrate that the deep network trained by ECG images provides outstanding classification performance of arrhythmic ECG signals and outperforms similar network architectures. Moreover, the proposed method has lower computational costs compared to existing methods and is more suitable for mobile device-based diagnosis systems as it does not involve any complex preprocessing process. Hence, the proposed approach provides a simple and robust automatic cardiac arrhythmia detection scheme for the classification of ECG arrhythmias.     

Detection and classification of diseased pine trees with different levels of severity from UAV remote sensing images

The accurate detection and classification of diseased pine trees with different levels of severity is important in terms of monitoring the growth of these trees and for preventing and controlling disease within pine forests. Our method combines a DDYOLOv5 with a ResNet50 network for detecting and classifying levels of pine tree disease from remote sensing UAV images. In this approach, images are preprocessed to increase the background diversity of the training samples, and efficient channel attention (ECA) and hybrid dilated convolution (HDC) modules are introduced to DDYOLOv5 to improve the detection accuracy. The ECA modules enable the network to focus on the characteristics of diseased pine trees, and solve the problem of low detection accuracy caused by the similarities in color and texture between diseased pine trees and the complex backgrounds. The HDC modules capture the contextual information of targets at different scales; they increase the receptive field to focus on targets of different sizes, and address the difficulty of detection caused by large variations in the shapes and sizes of diseased pine trees. In addition, a low confidence threshold is adopted to reduce missed detections and a ResNet50 classification network is applied to classify the detection results into different levels of severity, in order to reduce the number of false detections and improve the classification accuracy. Our experimental results show that the proposed method improves the precision by 13.55%, the recall by 5.06% and the F1-score by 9.71% on 8 test images compared with YOLOv5. Moreover, the detection and classification results from our approach show that it outperforms classical deep learning object detection methods such as Faster R-CNN and RetinaNet.     

Development of full‐field optical spatial coherence tomography system for automated identification of malaria using the multilevel ensemble classifier

Malaria is a life‐threatening infectious blood disease affecting humans and other animals caused by parasitic protozoans belonging to the Plasmodium type especially in developing countries. The gold standard method for the detection of malaria is through the microscopic method of chemically treated blood smears. We developed an automated optical spatial coherence tomographic system using a machine learning approach for a fast identification of malaria cells. In this study, 28 samples (15 healthy, 13 malaria infected stages of red blood cells) were imaged by the developed system and 13 features were extracted. We designed a multilevel ensemble‐based classifier for the quantitative prediction of different stages of the malaria cells. The proposed classifier was used by repeating k‐fold cross validation dataset and achieve a high‐average accuracy of 97.9% for identifying malaria infected late trophozoite stage of cells. Overall, our proposed system and multilevel ensemble model has a substantial quantifiable potential to detect the different stages of malaria infection without staining or expert.