Prediction Consistency Regularization for Learning with Noise Labels Based on Contrastive Clustering.

In the classification task, label noise has a significant impact on models' performance, primarily manifested in the disruption of prediction consistency, thereby reducing the classification accuracy. This work introduces a novel prediction consistency regularization that mitigates the impact of label noise on neural networks by imposing constraints on the prediction consistency of similar samples. However, determining which samples should be similar is a primary challenge. We formalize the similar sample identification as a clustering problem and employ twin contrastive clustering (TCC) to address this issue. To ensure similarity between samples within each cluster, we enhance TCC by adjusting clustering prior to distribution using label information. Based on the adjusted TCC's clustering results, we first construct the prototype for each cluster and then formulate a prototype-based regularization term to enhance prediction consistency for the prototype within each cluster and counteract the adverse effects of label noise. We conducted comprehensive experiments using benchmark datasets to evaluate the effectiveness of our method under various scenarios with different noise rates. The results explicitly demonstrate the enhancement in classification accuracy. Subsequent analytical experiments confirm that the proposed regularization term effectively mitigates noise and that the adjusted TCC enhances the quality of similar sample recognition.

Improvement of the YOLOv5 Model in the Optimization of the Brown Spot Disease Recognition Algorithm of Kidney Bean.

The kidney bean is an important cash crop whose growth and yield are severely affected by brown spot disease. Traditional target detection models cannot effectively screen out key features, resulting in model overfitting and weak generalization ability. In this study, a Bi-Directional Feature Pyramid Network (BiFPN) and Squeeze and Excitation (SE) module were added to a YOLOv5 model to improve the multi-scale feature fusion and key feature extraction abilities of the improved model. The results show that the BiFPN and SE modules show higher heat in the target location region and pay less attention to irrelevant environmental information in the non-target region. The detection Precision, Recall, and mean average Precision (mAP@0.5) of the improved YOLOv5 model are 94.7%, 88.2%, and 92.5%, respectively, which are 4.9% higher in Precision, 0.5% higher in Recall, and 25.6% higher in the mean average Precision compared to the original YOLOv5 model. Compared with the YOLOv5-SE, YOLOv5-BiFPN, FasterR-CNN, and EfficientDet models, detection Precision improved by 1.8%, 3.0%, 9.4%, and 9.5%, respectively. Moreover, the rate of missed and wrong detection in the improved YOLOv5 model is only 8.16%. Therefore, the YOLOv5-SE-BiFPN model can more effectively detect the brown spot area of kidney beans.

Low-dose imaging denoising with one pair of noisy images.

Low-dose imaging techniques have many important applications in diverse fields, from biological engineering to materials science. Samples can be protected from phototoxicity or radiation-induced damage using low-dose illumination. However, imaging under a low-dose condition is dominated by Poisson noise and additive Gaussian noise, which seriously affects the imaging quality, such as signal-to-noise ratio, contrast, and resolution. In this work, we demonstrate a low-dose imaging denoising method that incorporates the noise statistical model into a deep neural network. One pair of noisy images is used instead of clear target labels and the parameters of the network are optimized by the noise statistical model. The proposed method is evaluated using simulation data of the optical microscope, and scanning transmission electron microscope under different low-dose illumination conditions. In order to capture two noisy measurements of the same information in a dynamic process, we built an optical microscope that is capable of capturing a pair of images with independent and identically distributed noises in one shot. A biological dynamic process under low-dose condition imaging is performed and reconstructed with the proposed method. We experimentally demonstrate that the proposed method is effective on an optical microscope, fluorescence microscope, and scanning transmission electron microscope, and show that the reconstructed images are improved in terms of signal-to-noise ratio and spatial resolution. We believe that the proposed method could be applied to a wide range of low-dose imaging systems from biological to material science.

Non-destructive monitoring of maize LAI by fusing UAV spectral and textural features.

Leaf area index (LAI) is an essential indicator for crop growth monitoring and yield prediction. Real-time, non-destructive, and accurate monitoring of crop LAI is of great significance for intelligent decision-making on crop fertilization, irrigation, as well as for predicting and warning grain productivity. This study aims to investigate the feasibility of using spectral and texture features from unmanned aerial vehicle (UAV) multispectral imagery combined with machine learning modeling methods to achieve maize LAI estimation. In this study, remote sensing monitoring of maize LAI was carried out based on a UAV high-throughput phenotyping platform using different varieties of maize as the research target. Firstly, the spectral parameters and texture features were extracted from the UAV multispectral images, and the Normalized Difference Texture Index (NDTI), Difference Texture Index (DTI) and Ratio Texture Index (RTI) were constructed by linear calculation of texture features. Then, the correlation between LAI and spectral parameters, texture features and texture indices were analyzed, and the image features with strong correlation were screened out. Finally, combined with machine learning method, LAI estimation models of different types of input variables were constructed, and the effect of image features combination on LAI estimation was evaluated. The results revealed that the vegetation indices based on the red (650 nm), red-edge (705 nm) and NIR (842 nm) bands had high correlation coefficients with LAI. The correlation between the linearly transformed texture features and LAI was significantly improved. Besides, machine learning models combining spectral and texture features have the best performance. Support Vector Machine (SVM) models of vegetation and texture indices are the best in terms of fit, stability and estimation accuracy (R2 = 0.813, RMSE = 0.297, RPD = 2.084). The results of this study were conducive to improving the efficiency of maize variety selection and provide some reference for UAV high-throughput phenotyping technology for fine crop management at the field plot scale. The results give evidence of the breeding efficiency of maize varieties and provide a certain reference for UAV high-throughput phenotypic technology in crop management at the field scale.

Demonstration of programmable light intensity of a micro-LED with a Hf-based ferroelectric ITZO TFT for Mura-free displays.

We demonstrate the programmable light intensity of a micro-LED by compensating threshold voltage variability of thin-film transistors (TFTs) by introducing a non-volatile programmable ferroelectric material, HfZrO2 (HZO) into the gate stack of the TFT. We fabricated an amorphous ITZO TFT, ferroelectric TFTs (FeTFTs), and micro-LEDs and verified the feasibility of our proposed current-driving active matrix circuit. Importantly, we successfully present the programmed multi-level lighting of the micro-LED, utilizing partial polarization switching in the a-ITZO FeTFT. We expect that this approach will be highly promising for the next-generation display technology, replacing complicated threshold voltage compensation circuits with a simple a-ITZO FeTFT.

Predicting willingness to donate blood based on machine learning: two blood donor recruitments during COVID-19 outbreaks.

Machine learning methods are a novel way to predict and rank donors' willingness to donate blood and to achieve precision recruitment, which can improve the recruitment efficiency and meet the challenge of blood shortage. We collected information about experienced blood donors via short message service (SMS) recruitment and developed 7 machine learning-based recruitment models using PyCharm-Python Environment and 13 features which were described as a method for ranking and predicting donors' intentions to donate blood with a floating number between 0 and 1. Performance of the prediction models was assessed by the Area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score in the full dataset, and by the accuracy in the four sub-datasets. The developed models were applied to prospective validations of recruiting experienced blood donors during two COVID-19 pandemics, while the routine method was used as a control. Overall, a total of 95,476 recruitments via SMS and their donation results were enrolled in our modelling study. The strongest predictor features for the donation of experienced donors were blood donation interval, age, and donation frequency. Among the seven baseline models, the eXtreme Gradient Boosting (XGBoost) and Support vector machine models (SVM) achieved the best performance: mean (95%CI) with the highest AUC: 0.809 (0.806-0.811), accuracy: 0.815 (0.812-0.818), precision: 0.840 (0.835-0.845), and F1 score of XGBoost: 0.843 (0.840-0.845) and recall of SVM: 0.991 (0.988-0.994). The hit rate of the XGBoost model alone and the combined XGBoost and SVM models were 1.25 and 1.80 times higher than that of the conventional method as a control in 2 recruitments respectively, and the hit rate of the high willingness to donate group was 1.96 times higher than that of the low willingness to donate group. Our results suggested that the machine learning models could predict and determine the experienced donors with a strong willingness to donate blood by a ranking score based on personalized donation data and demographical details, significantly improve the recruitment rate of blood donors and help blood agencies to maintain the blood supply in emergencies.

Cryptanalysis of an optical cryptosystem with uncertainty quantification in a probabilistic model.

In this paper, a modified probabilistic deep learning method is proposed to attack the double random phase encryption by modeling the conditional distribution of plaintext. The well-trained probabilistic model gives both predictions of plaintext and uncertainty quantification, the latter of which is first introduced to optical cryptanalysis. Predictions of the model are close to real plaintexts, showing the success of the proposed model. Uncertainty quantification reveals the level of reliability of each pixel in the prediction of plaintext without ground truth. Subsequent simulation experiments demonstrate that uncertainty quantification can effectively identify poor-quality predictions to avoid the risk of unreliability from deep learning models.

Natural speckle-based watermarking with random-like illuminated decoding.

We propose an optical watermarking method based on a natural speckle pattern. In the watermarking process, the watermark information is embedded into the natural speckle pattern. Then the random-like watermarked image is generated with the proposed grayscale reordering algorithm. During the extraction procedure, the watermarked image is projected to the natural speckle pattern as illumination. Subsequently, they are incoherently superimposed to extract the watermark information directly by human vision. Optical experiments and a hypothesis test are conducted to demonstrate the proposed method with high reliability, imperceptibility and robustness. The proposed method is the first watermarking method utilizing the natural diffuser as the core element in encoding and decoding.