Classification of foodborne bacteria using hyperspectral microscope imaging technology coupled with convolutional neural networks‡.

Foodborne pathogens have become ongoing threats in the food industry, whereas their rapid detection and classification at an early stage are still challenging. To address early and rapid detection, hyperspectral microscope imaging (HMI) technology combined with convolutional neural networks (CNN) was proposed to classify foodborne bacterial species at the cellular level. HMI technology can simultaneously obtain both spatial and spectral information of different live bacterial cells, while two CNN frameworks, U-Net and one-dimensional CNN (1D-CNN), were employed to accelerate the data analysis process. U-Net was used for automating cellular regions of interest (ROI) segmentation, which generated accurate cell-ROI masks in a shorter timeframe than the conventional Otsu or Watershed methods. The 1D-CNN was employed for classifying the spectral profiles extracted from cell-ROI and resulted in a higher accuracy (90%) than k-nearest neighbor (81%) and support vector machine (81%). Overall, the CNN-assisted HMI technology showed potential for foodborne bacteria detection.

Assessment of Tomato Maturity in Different Layers by Spatially Resolved Spectroscopy.

Tomato maturity is important to determine the fruit shelf life and eating quality. The objective of this research was to evaluate tomato maturity in different layers by using a newly developed spatially resolved spectroscopic system over the spectral region of 550-1650 nm. Thirty spatially resolved spectra were obtained for 600 tomatoes, 100 for each of the six maturity stages (i.e., green, breaker, turning, pink, light red, and red). Support vector machine discriminant analysis (SVMDA) models were first developed for each of individual spatially resolved (SR) spectra to compare the classification results of two sides. The mean spectra of two sides with the same source-detector distances were employed to determine the model performance of different layers. SR combination by averaging all the SR spectra was also subject to comparison with the classification model performance. The results showed large source-detector distances would be helpful for evaluating tomato maturity, and the mean_SR 15 obtained excellent classification results with the total classification accuracy of 98.3%. Moreover, the classification results were distinct for two sides of the probe, which demonstrated even if in the same source-detector distances, the classification results were influenced by the measurement location due to the heterogeneity for tomato. The mean of all SR spectra could only improve the classification results based on the first three mean_SR spectra, but could not obtain the accuracy as good as the following mean_SR spectra. This study demonstrated that spatially resolved spectroscopy has potential for assessing tomato maturity in different layers.

Identification of Apple Varieties Using a Multichannel Hyperspectral Imaging System.

This paper reports the nondestructive detection of apple varieties using a multichannel hyperspectral imaging system consisting of an illumination fiber and 30 detection fibers arranged at source-detector distances of 1.5-36 mm over the spectral range of 550-1650 nm. Spatially resolved (SR) spectra were obtained for 1500 apples, 500 each of three varieties from the same orchard to avoid environmental and geographical influences. Partial least squares discriminant analysis (PLSDA) models were developed for single SR spectra and spectral combinations to compare their performance of variety detection. To evaluate the effect of spectral range on variety detection, three types of spectra (i.e., visible region: 550-780 nm, near-infrared region: 780-1650 nm, full region: 550-1650 nm) were analyzed and compared. The results showed that the single SR spectra presented a different accuracy for apple variety classification, and the optimal SR spectra varied with spectral types. Spectral combinations had better accuracies for variety detection with best overall classifications of 99.4% for both spectral ranges in the NIR and full regions; however, the spectral combination could not improve the results over the optimal single SR spectra in the visible region. Moreover, the recognition of golden delicious (GD) was better than those of the other two varieties, with the best classification accuracy of 100% for three types of spectra. Overall, the multichannel hyperspectral imaging system provides more spatial-spectral information for the apples, and the results demonstrate that the technique gave excellent classifications, which suggests that the multichannel hyperspectral imaging system has potential for apple variety detection.

An investigation on the relationship among marbling features, physiological age and Warner-Bratzler Shear force of steer longissimus dorsi muscle.

Researchers nowadays have paid much attention to the relationships between tenderness and marbling, or physiological age. While the marbling was mainly evaluated qualitatively with scores or grades, and rarely related with physiological age. Present study was carried out to analyze the marbling features of longissimus dorsi muscle between the 12th and 13th ribs from 18, 36, 54 and 72 months old Simmental steers were quantitatively described with area and perimeter using computer vision technique. Relationship between Warner-Bratzler Shear force (WBSF), physiological age and the marbling features were examined performing regression analysis. The results revealed that WBSF positively correlated with physiological age, but negatively with marbling area and perimeter. Regression analysis showed that the relationship between the shear force and the steers' age was more close to the quadratic curve (R2 = 0.996) and exponential curve (R2 = 0.957). It was observed during study that marbling grew with steers age. Marbling features were in linear correlation with the steers' age, with R2 = 0.927 for marbling area and R2 = 0.935 for marbling perimeter. The industries in future may speculate beef tenderness and physiological age based on the marbling features (area and perimeter), which can be determined through the online image acquisition system and image processing.

Detection of diluted contaminants on chicken carcasses using a two-dimensional scatter plot based on a two-dimensional hyperspectral correlation spectrum.

A two-dimensional (2D) scatter plot method based on the 2D hyperspectral correlation spectrum is proposed to detect diluted blood, bile, and feces from the cecum and duodenum on chicken carcasses. First, from the collected hyperspectral data, a set of uncontaminated regions of interest (ROIs) and four sets of contaminated ROIs were selected, whose average spectra were treated as the original spectrum and influenced spectra, respectively. Then, the difference spectra were obtained and used to conduct correlation analysis, from which the 2D hyperspectral correlation spectrum was constructed using the analogy method of 2D IR correlation spectroscopy. Two maximum auto-peaks and a pair of cross peaks appeared at 656 and 474 nm. Therefore, 656 and 474 nm were selected as the characteristic bands because they were most sensitive to the spectral change induced by the contaminants. The 2D scatter plots of the contaminants, clean skin, and background in the 474- and 656-nm space were used to distinguish the contaminants from the clean skin and background. The threshold values of the 474- and 656-nm bands were determined by receiver operating characteristic (ROC) analysis. According to the ROC results, a pixel whose relative reflectance at 656 nm was greater than 0.5 and relative reflectance at 474 nm was lower than 0.3 was judged as a contaminated pixel. A region with more than 50 pixels identified was marked in the detection graph. This detection method achieved a recognition rate of up to 95.03% at the region level and 31.84% at the pixel level. The false-positive rate was only 0.82% at the pixel level. The results of this study confirm that the 2D scatter plot method based on the 2D hyperspectral correlation spectrum is an effective method for detecting diluted contaminants on chicken carcasses.

The use of steam explosion to increase the nutrition available from rice straw.

In the present study, rice straw was pretreated using steam-explosion (ST) technique to improve the enzymatic hydrolysis of potential reducing sugars for feed utilization. The response surface methodology based on central composite design was used to optimize the effects of steam pressure, pressure retention time, and straw moisture content on the yield of reducing sugar. All the investigated variables had significant effects (P < 0.001) on the reducing sugar yield. The optimum yield of 30.86% was obtained under the following pretreatment conditions: steam pressure, 1.54 MPa; pressure retention time, 140.5 Sec; and straw moisture content, 41.6%. The yield after thermal treatment under the same conditions was approximately 16%. Infrared (IR) radiation analysis showed a decrease in the cellulose IR crystallization index. ST noticeably increases reducing sugars in rice straw, and this technique may also be applicable to other cellulose/lignin sources of biomass.

[A dual-band algorithm to detect contaminants with low visibility on chicken carcass surface].

A novel dual-band algorithm for detecting contaminants with low visibility on chicken carcass surface based on hyperspectral image was proposed. Firstly, The 675 nm band image, in which the identity of the intensity within ROI (Region of Interest) is the best and the spectrum difference between ROI and the edge of the ROI is the biggest, was chosen from the hyperspectral data for binarization and the mask was extracted by using region growing on the biggest connected area. Then the "and" operation between the mask and the 400 nm band image with the largest discriminability of contaminants was carried out. The max ROI which can self adapt according to the position and shape of the chicken carcass was obtained. Finally, the labeling method was used to recognize if there are contaminations within the segmented ROI. The results showed that through the proposed method, the max ROIs which could self adapt to the position and shape of the chicken carcass were extracted and the average size of the ROI was bigger than 176% compared to that by existing methods. The average correct identification rate of contaminations such as blood, bile and feces was 81.6%.

Parameter calibration of discrete element model for gluten densification molding.

The densified powder material is convenient for storage and transportation, with broad market application prospects. In this study, the discrete element model parameters required for simulating gluten densification were calibrated using the Hertz-Mindlin with JKR contact model. Initially, physical testing techniques were utilized to assess the size distribution, density, and angle of repose (AoR) of gluten particles. Following this, the Plackett-Burman test, the steepest ascent test, and the Box-Behnken test were conducted, and the significant factors were obtained: The coefficient of rolling friction (P-P) was 1.038, the coefficient of static friction (P-P) was 0.071, and the surface energy (P-P) was 0.047. Finally, the AoR and densification simulations were performed under the optimal parameter combination, along with validation tests. The results showed that the relative error between the simulated and tested AoR was 0.52%. The compression ratio and compression force curves of simulated and actual were similar.

Measurement and calibration of gluten pellets discrete element parameters at varied moisture content.

To quickly calibrate the discrete element parameters (DEP) of pellets with different moisture content (MC), the angle of repose (AoR) was taken as the target value to conduct experimental and simulation research on gluten pellets. The experimental method obtained the intrinsic parameters, contact parameters, and AoR of pellets with different moisture content. The parameters differed significantly under different moisture content (p < 0.05). The AoR-MC model (R2  = 0.987) was established. The Plackett-Burman test, steepest ascent test, and center compound test were carried out to establish the AoR-DEP model (R2  = 0.969) with a relative error less than or equal to 2.07%. The MC-DEP model was derived, and verified by the side plate lifting method with a relative error less than or equal to 2.58%. This paper provides a new method for calibrating DEP under different moisture content.

Compression Characteristics and Fracture Simulation of Gluten Pellet.

Gluten pellets are readily broken on packaging and transportation. This study aimed to research mechanical properties (elastic modulus, compressive strength, failure energy) with different moisture contents and aspect ratios under different compressive directions. The mechanical properties were examined with a texture analyzer. The results revealed that the material properties of the gluten pellet are anisotropic, and it was more likely to cause crushing during radial compression. The mechanical properties were positively correlated with the moisture content. The aspect ratio had no significant effect (p > 0.05) on the compressive strength. The statistical function model (p < 0.01; R2 ≥ 0.774) for mechanical properties and moisture content fitted well with the test data. The minimum elastic modulus, compressive strength, and failure energy of standards-compliant pellets (with moisture content less than 12.5% d.b.) were 340.65 MPa, 6.25 MPa, and 64.77 mJ, respectively. Moreover, a finite element model with cohesive elements was established using Abaqus software (Version 2020, Dassault Systèmes, Paris, France) to simulate the compression rupture form of gluten pellets. The relative error of the fracture stress in the axial and radial directions between the simulation results and the experimental value was within 4-7%.

Tea Chrysanthemum Detection by Leveraging Generative Adversarial Networks and Edge Computing.

A high resolution dataset is one of the prerequisites for tea chrysanthemum detection with deep learning algorithms. This is crucial for further developing a selective chrysanthemum harvesting robot. However, generating high resolution datasets of the tea chrysanthemum with complex unstructured environments is a challenge. In this context, we propose a novel tea chrysanthemum - generative adversarial network (TC-GAN) that attempts to deal with this challenge. First, we designed a non-linear mapping network for untangling the features of the underlying code. Then, a customized regularization method was used to provide fine-grained control over the image details. Finally, a gradient diversion design with multi-scale feature extraction capability was adopted to optimize the training process. The proposed TC-GAN was compared with 12 state-of-the-art generative adversarial networks, showing that an optimal average precision (AP) of 90.09% was achieved with the generated images (512 × 512) on the developed TC-YOLO object detection model under the NVIDIA Tesla P100 GPU environment. Moreover, the detection model was deployed into the embedded NVIDIA Jetson TX2 platform with 0.1 s inference time, and this edge computing device could be further developed into a perception system for selective chrysanthemum picking robots in the future.

Medicinal Chrysanthemum Detection under Complex Environments Using the MC-LCNN Model.

Medicinal chrysanthemum detection is one of the desirable tasks of selective chrysanthemum harvesting robots. However, it is challenging to achieve accurate detection in real time under complex unstructured field environments. In this context, we propose a novel lightweight convolutional neural network for medicinal chrysanthemum detection (MC-LCNN). First, in the backbone and neck components, we employed the proposed residual structures MC-ResNetv1 and MC-ResNetv2 as the main network and embedded the custom feature extraction module and feature fusion module to guide the gradient flow. Moreover, across the network, we used a custom loss function to improve the precision of the proposed model. The results showed that under the NVIDIA Tesla V100 GPU environment, the inference speed could reach 109.28 FPS per image (416 × 416), and the detection precision (AP50) could reach 93.06%. Not only that, we embedded the MC-LCNN model into the edge computing device NVIDIA Jetson TX2 for real-time object detection, adopting a CPU-GPU multithreaded pipeline design to improve the inference speed by 2FPS. This model could be further developed into a perception system for selective harvesting chrysanthemum robots in the future.

Pseudo high-frequency boosts the generalization of a convolutional neural network for cassava disease detection.

Frequency is essential in signal transmission, especially in convolutional neural networks. It is vital to maintain the signal frequency in the neural network to maintain the performance of a convolutional neural network. Due to destructive signal transmission in convolutional neural network, signal frequency downconversion in channels results into incomplete spatial information. In communication theory, the number of Fourier series coefficients determines the integrity of the information transmitted in channels. Consequently, the number of Fourier series coefficients of the signals can be replenished to reduce the information transmission loss. To achieve this, the ArsenicNetPlus neural network was proposed for signal transmission modulation in detecting cassava diseases. First, multiattention was used to maintain the long-term dependency of the features of cassava diseases. Afterward, depthwise convolution was implemented to remove aliasing signals and downconvert before the sampling operation. Instance batch normalization algorithm was utilized to keep features in an appropriate form in the convolutional neural network channels. Finally, the ArsenicPlus block was implemented to generate pseudo high-frequency in the residual structure. The proposed method was tested on the Cassava Datasets and compared with the V2-ResNet-101, EfficientNet-B5, RepVGG-B3g4 and AlexNet. The results showed that the proposed method performed [Formula: see text] in terms of accuracy, 1.2440 in terms of loss, and [Formula: see text] in terms of the F1-score, outperforming the comparison algorithms.

[Research on identification of cucumber, stem and leaf based on spectrum analysis technology].

To be able to quickly identify the cucumber real time, the present paper studied the near infrared reflectance characteristics of cucumber, stem and leaf. Spectral reflectance of 138 samples (46 cucumbers, 46 stems and 46 leaves) was collected using near infrared spectroscopy in the band range of 600-1 099 nm indoor. After Savitzky-Golay smoothing preprocessing, random 108 spectral samples were put forward as calibration set. The weighted deviation method was used for choosing the spectral bands 690-950 nm that include much more information. The samples were analyzed by PCA method to extract the principal component scores, combining the Mahalanobis distance method the recognition model was established, and seven abnormal samples were excluded. The partial least squares (PLS) model was established by remaining 101 samples spectra of calibration set, which was used for predicting the validation set (30 samples except of the calibration set). The result shows that the correlation of the predicted value and the actual value reaches up to 0.994 1, and the correct recognition rate is 100%. This significantly illustrates that the near infrared spectral reflectance characteristics are different among the cucumbers, stems and leaves, which can be successfully applied to recognition of cucumber by the method. The developed technique can provide a new method for cucumber identification.

Identifying non-O157 Shiga toxin-producing Escherichia coli (STEC) using deep learning methods with hyperspectral microscope images.

Non-O157 Shiga toxin-producing Escherichia coli (STEC) serogroups such as O26, O45, O103, O111, O121 and O145 often cause illness to people in the United States and the conventional identification of these "Big-Six" are complex. The label-free hyperspectral microscope imaging (HMI) method, which provides spectral "fingerprints" information of bacterial cells, was employed to classify serogroups at the cellular level. In spectral analysis, principal component analysis (PCA) method and stacked auto-encoder (SAE) method were conducted to extract principal spectral features for classification task. Based on these features, multiple classifiers including linear discriminant analysis (LDA), support vector machine (SVM) and soft-max regression (SR) methods were evaluated. Different sizes of datasets were also tested in search for the suitable classification models. Among the results, SAE-based classification models performed better than PCA-based models, achieving classification accuracy of SAE-LDA (93.5%), SAE-SVM (94.9%) and SAE-SR (94.6%), respectively. In contrast, classification results of PCA-based methods such as PCA-LDA, PCA-SVM and PCA-SR were only 75.5%, 85.7% and 77.1%, respectively. The results also suggested the increasing number of training samples have positive effects on classification models. Taking advantage of increasing dataset, the SAE-SR classification model finally performed better than others with average accuracy of 94.9% in classifying STEC serogroups. Specifically, O103 serogroup was classified with the highest accuracy of 97.4%, followed by O111 (96.5%), O26 (95.3%), O121 (95%), O145 (92.9%) and O45 (92.4%), respectively. Thus, the HMI technology coupled with SAE-SR classification model has the potential for "Big-Six" identification.

Broiler stunned state detection based on an improved fast region-based convolutional neural network algorithm.

An improved fast region-based convolutional neural network (RCNN) algorithm is proposed to improve the accuracy and efficiency of recognizing broilers in a stunned state. The algorithm recognizes 3 stunned state conditions: insufficiently stunned, moderately stunned, and excessively stunned. Image samples of stunned broilers were collected from a slaughter line using an image acquisition platform. According to the format of PASCAL VOC (pattern analysis, statistical modeling, and computational learning visual object classes) dataset, a dataset for each broiler stunned state condition was obtained using an annotation tool to mark the chicken head and wing area in the original image. A rotation and flip data augmentation method was used to enhance the effectiveness of the datasets. Based on the principle of a residual network, a multi-layer residual module (MRM) was constructed to facilitate more detailed feature extraction. A model was then developed (entitled here Faster-RCNN+MRMnet) and used to detect broiler stunned state conditions. When applied to a reinforcing dataset containing 27,828 images of chickens in a stunned state, the identification accuracy of the model was 98.06%. This was significantly higher than both the established back propagation neural network model (90.11%) and another Faster-RCNN model (96.86%). The proposed algorithm can complete the inspection of the stunned state of more than 40,000 broilers per hour. The approach can be used for online inspection applications to increase efficiency, reduce labor and cost, and yield significant benefits for poultry processing plants.

Efficient pyrolysis of ginkgo biloba leaf residue and pharmaceutical sludge (mixture) with high production of clean energy: Process optimization by particle swarm optimization and gradient boosting decision tree algorithm.

Production of sustainable clean energy can be achieved by co-pyrolysis of agricultural residues and wastewater sludge. Herein, non-additive thermal behaviour of co-pyrolysis of pharmaceutical sludge and ginkgo biloba leaf residues was investigated. Synergistic effect of co-pyrolysis was not obvious at elevated temperatures. Further, kinetics of co-pyrolysis was studied by fitting Coats-Redfern integration method to thermogravimetric (TG) curve. The change of heat and mass transfer in the reactor caused the change of dynamic parameters. Moreover, hybrid particle swarm optimization and gradient boosting decision tree (PSO-GBDT) algorithm was designed to boost the energy production at full-scale pyrolysis plant by monitoring TG curves. PSO-GBDT model well predicts mass loss rate of the mixture at different heating rates confirming that co-pyrolysis of PS and GBLR can results in high energy production by increasing PS pyrolysis. Designing PSO-GBDT model help to reduced waste production by resourceful treatment of waste in to energy.