Interpretable deep learning by disentangling features using modified Siamese network

This dissertation explores the critical role of loss functions in enhancing the predictive performance of deep machine learning models. Loss functions are an integral element of all the ongoing advances we witness daily in this domain. I design custom loss functions and their impacts on various machine learning tasks, particularly in computer vision. In the first stage of my research, I aim to improve the prediction performance of deep learning models by providing them with more precise feedback associated with task requirements. This led me to create the concept of assistive loss functions. My first proposed loss function, inspired by the Active Shape Model (ASM), is ASM Loss, designed to improve the detection of facial landmarks and head pose accuracy in facial images. Additionally, influenced by Knowledge Distillation (KD), I introduce KD-Loss, a loss designed specifically for regression tasks, focusing on detecting facial landmarks. Proceeding to the second phase of my research, I focus on regression problems. Building upon my prior work, I introduce Adaptive Coordinate-based Regression (ACR)-Loss, a context-aware loss function with face alignment implication. I provide a mechanism to dynamically measure how hard the localization of each landmark point is, and accordingly, ACR-Loss provides unique feedback to the model. Furthermore, I introduce Intensity-aware Categorical (IC)-Loss for detecting sagittal cervical spine landmark points in medical images. It effectively deploys characteristics from classification problems to enhance regression optimization. Both ACR-Loss and IC-Loss comprise contextual information, leading to a significant enhancement in localization precision. In the third phase, I turned my attention toward classification problems within the domain of deep machine learning. I introduce Adaptive Correlation-based (Ad-corre) Loss, designed to enhance the discriminative capabilities of deep learning models, with a specific focus on improving classification accuracy in automatic facial expression recognition. Furthermore, I introduce Informative (Info) Loss, a specialized loss function developed for medical applications, specifically aimed at distinguishing between Mild Cognitive Impairment and Normal Cognition groups based on transcriptions of interviews conducted as part of the I-CONECT study. Overall, I have proposed custom loss functions to improve the predictive performance in both regression and classification problems. Furthermore, I have introduced assistive loss functions, designed to enhance the performance of the main loss functions. The collective outcomes of this extensive exploration highlight the critical role of loss functions, adapting and optimizing them for various domains in deep machine learning. This research suggests and proposes a variety of methodologies for designing adaptive and context-aware loss functions within deep machine learning.

Face recognition (FR) is a method that uses face feature analysis and comparison to identify or verify individuals. Siamese neural networks (SNNs) are an effective method for FR, providing high accuracy and versatility, especially in situations where data is restricted. Unlike standard neural networks, SNNs learn to distinguish between pairs of inputs rather than individual inputs. However, detecting and recognizing faces in unconstrained environments poses a significant challenge due to various factors such as head pose, illumination, and facial expression variations. The aim of this paper is to design and develop an efficient approach based on SNNs and Transfer Learning methods. For this purpose LFW dataset and transfer learning architectures like VGG-16, EfficientNet, RestNet50 and ConvNext have been utilised. Performance of the architectures were measured using 5-Fold cross validation. According to results, EfficientNet, RestNet50 and ConvNext produced 78% accuracy, 95% and 93 % accuracy respectively. SNN with VGG-16 exhibited a low loss and produced the best accuracy in face recognition with 96%.

Potatoes are one of the horticultural commodities with significant trade value both domestically and internationally. To produce high-quality potatoes, healthy and disease-free potato plants are essential. The most common diseases affecting potato plants are late blight and early blight. These diseases appear randomly in different positions and sizes on potato leaves, resulting in numerous combinations of infected leaves. This study proposes an architecture focused on a similarity-based approach, namely the Siamese Neural Network (SNN). SNN can recognize images by comparing two or more images and categorizing the test image accordingly. Thus, SNN has an advantage over classification-based approaches as it can identify various combinations of disease spots on potato plants using a similarity-based approach. This study is divided into two main scenarios: testing with data categories which were previously seen during the training process (traditional testing) and testing with the addition of new data categories that were not seen during training. In the first scenario, SNN showed better accuracy with an accuracy rate of 98.4%, while in the second scenario, SNN achieved an accuracy of 97.1%. That result suggests that SNN can categorize data very well, even recognizing data which never seen during training. These results offer hope that SNN can recognize more disease spots/patterns on potato plants or even identify new diseases by adding these new diseases to the SNN support set without retraining.

Plant fungal diseases are one of the most important causes of crop yield losses. Therefore, plant disease identification algorithms have been seen as a useful tool to detect them at early stages to mitigate their effects. Although deep-learning based algorithms can achieve high detection accuracies, they require large and manually annotated image datasets that is not always accessible, specially for rare and new diseases. This study focuses on the development of a plant disease detection algorithm and strategy requiring few plant images (Few-shot learning algorithm). We extend previous work by using a novel challenging dataset containing more than 100,000 images. This dataset includes images of leaves, panicles and stems of five different crops (barley, corn, rape seed, rice, and wheat) for a total of 17 different diseases, where each disease is shown at different disease stages. In this study, we propose a deep metric learning based method to extract latent space representations from plant diseases with just few images by means of a Siamese network and triplet loss function. This enhances previous methods that require a support dataset containing a high number of annotated images to perform metric learning and few-shot classification. The proposed method was compared over a traditional network that was trained with the cross-entropy loss function. Exhaustive experiments have been performed for validating and measuring the benefits of metric learning techniques over classical methods. Results show that the features extracted by the metric learning based approach present better discriminative and clustering properties. Davis-Bouldin index and Silhouette score values have shown that triplet loss network improves the clustering properties with respect to the categorical-cross entropy loss. Overall, triplet loss approach improves the DB index value by 22.7% and Silhouette score value by 166.7% compared to the categorical cross-entropy loss model. Moreover, the F-score parameter obtained from the Siamese network with the triplet loss performs better than classical approaches when there are few images for training, obtaining a 6% improvement in the F-score mean value. Siamese networks with triplet loss have improved the ability to learn different plant diseases using few images of each class. These networks based on metric learning techniques improve clustering and classification results over traditional categorical cross-entropy loss networks for plant disease identification.

Siamese neural network is a very powerful architecture for both feature extraction and metric learning. It usually consists of several networks that share weights. The Siamese concept is topology-agnostic and can use any neural network as its backbone. The two most popular loss functions for training these networks are the triplet and contrastive loss functions. In this paper, we propose two novel loss functions, named Fisher Discriminant Triplet (FDT) and Fisher Discriminant Contrastive (FDC). The former uses anchor-neighbor-distant triplets while the latter utilizes pairs of anchor-neighbor and anchor-distant samples. The FDT and FDC loss functions are designed based on the statistical formulation of the Fisher Discriminant Analysis (FDA), which is a linear subspace learning method. Our experiments on the MNIST and two challenging and publicly available histopathology datasets show the effectiveness of the proposed loss functions.

Handwritten signature verification is a critical aspect of personal authentication in numerous sectors, including banking, legal, and access control systems. This paper provides a comprehensive investigation into writer-dependent signature verification techniques, spanning from traditional approaches such as Histogram of Oriented Gradients (HOG) and Scale-Invariant Feature Transform (SIFT) to cutting-edge deep learning architectures like Siamese and triplet networks. Through extensive experimentation and meticulous analysis conducted on benchmark datasets, we rigorously compare and evaluate the efficacy of these approaches. Our study reveals intriguing insights into the performance characteristics of different methodologies, shedding light on their strengths, weaknesses, and areas for improvement. Notably, we observe that while traditional feature-based methods excel in capturing discriminative information from signature images, deep learning architectures offer enhanced flexibility and adaptability, particularly in scenarios with large and diverse datasets. By combining the strengths of both paradigms, we demonstrate the potential for building robust and reliable signature verification systems capable of handling real-world challenges and variations. We highlight the importance of carefully curated datasets and well-designed network architectures in achieving optimal results, thereby providing valuable guidelines for practitioners and researchers in the field. In conclusion, this study contributes significant insights to the advancement of handwritten signature verification systems, emphasizing the importance of leveraging both traditional methodologies and modern deep learning techniques for building secure and dependable authentication mechanisms. By addressing key challenges and exploring novel approaches, our research aims to foster innovation and enhance the security measures associated with handwritten signature authentication in diverse applications. Keywords—Handwritten signature verification, Writer-dependent approach, Feature-based methods, Deep learning, Siamese networks, Triplet networks, Benchmark datasets, Performance evaluation, Authentication mechanisms, Security measures.

Accurate disease recognition from wheat plants is essential to mitigate the effects and to stop the spread of the diseases. State-of-the-art algorithms have developed through deep learning which recognizes the disease from field images. Although such methods produce highly accurate results, they require tons of labeled images, and work only on the classes which are involved in the training phase. Thus, to get the recognition against the different classes, the model is required to retrain. This article has proposed a wheat disease recognition network based on one-shot learning which not only needs a small number of images for training, but also can accommodate new categories because it can be trained even on few images of a new type. So, farmers can retrain the network just by giving a few images of plants containing the concerned disease, and start testing immediately. We have used the MobileNetv3 network as a feature extractor which is extremely fast and accurate classification. This network is fine-tuned on the PlantVillage dataset, while the last two dense layers are fine-tuned on the plant images of 11 wheat disease, those images are taken from the CGIAR Crop Disease dataset and Google images. The whole One-shot network is trained on 440 images having 40 images of each class. Siamese networks are used for producing the encodings of input images, then the absolute difference is calculated between encodings, and similarity scores are determined through the Sigmoid unit. It assigns 1 score to similar images, while 0 to dissimilar images. Our Mobilenetv3 model has achieved around 98% training and 96% validation, while the whole one-shot network has achieved more than 92% accuracy, 84% precision, and 85 recall. Our proposed system requires just a few images of a new type for training, instead of retraining the network as in standard classification networks.

In the last few years, mechanical fault diagnosis (FD) based on deep learning has been systematically applied to various industrial fields. However, most methods rely on enough labeled data, but collecting enough sample of failure is timeconsuming and labor-intensive. It remains challenging to train a model for FD with numbered training data and to function correctly under intricate operating scenario. In response to this problem, the FD method based on Siamese neural network (SNN) has shown promising results in recent years, especially in the case of insufficient training samples; the fault diagnosis work has achieved good results and has been widely concerned by researchers at domestic and foreign. To better combine the SNN for fault diagnosis and give full play to the advantages of the SNN in fault diagnosis, this paper summarizes and analyzes the fault diagnosis methods based on SNN. Firstly, the basic framework of the SNN method is introduced, and the advantages and application scenarios of the framework are pointed out. Secondly, the existing fault diagnosis methods based on SNN are classified, including the principle of the technique, improved methods, and shortcomings. Finally, the FD methods based on SNN are summarized, and the development of the SNN in FD is in prospect.

An empirical study of the impact of masks on face recognition

Deep learning is a very active research field in machine learning. The most popular used convolutional neural networks (CNN) is one of deep learning models. However., deep learning suffers from a time consuming training process because of a large number of parameters in filters and layers to be learned. Broad Learning System (BLS) offers an alternative way of machine learning in deep structure. BLS is established based on the idea of the random vector function-link neural network (RVFLNN) which eliminates the drawback of long training process and also provides the generalization capability in function approximation. In this paper, a CNN-based broad learning system (CNNBLS) which is more useful than BLS for computer vision is proposed. In the proposed method, convolution and max pooling are used for feature extraction and then the principal component analysis (PCA) is used for feature dimensionality reduction, following a ridge regression learning for pattern classification. Parameters in convolution filters are randomly initialized which is similar with BLS and RVFLNN. Therefore the proposed method needs less parameters to be optimized at the same time it adopts the very useful feature extraction method for computer vision. Experimental results on the MNIST handwritten digits recognition data set and the NYU NORB object recognition data set demonstrate the effectiveness of the proposed method.

A novel feature extraction method that utilizes nonlinear mapping from the original data space to the feature space is presented in this dissertation. Feature extraction methods aim to find compact representations of data that are easy to classify. Measurements with similar values are grouped to same category, while those with differing values are deemed to be of separate categories. For most practical systems, the meaningful features of a pattern class lie in a low dimensional nonlinear constraint region (manifold) within the high dimensional data space. A learning algorithm to model this nonlinear region and to project patterns to this feature space is developed. Least squares estimation approach that utilizes interdependency between points in training patterns is used to form the nonlinear region. The proposed feature extraction strategy is employed to improve face recognition accuracy under varying illumination conditions and facial expressions. Though the face features show variations under these conditions, the features of one individual tend to cluster together and can be considered as a neighborhood. Low dimensional representations of face patterns in the feature space may lie in a nonlinear constraint region, which when modeled leads to efficient pattern classification. A feature space encompassing multiple pattern classes can be trained by modeling a separate constraint region for each pattern class and obtaining a mean constraint region by averaging all the individual regions. Unlike most other nonlinear techniques, the proposed method provides an easy intuitive way to place new points onto a nonlinear region in the feature space. The proposed feature extraction and classification method results in improved accuracy when compared to the classical linear representations. Face recognition accuracy is further improved by introducing the concepts of modularity, discriminant analysis and phase congruency into the proposed method. In the modular approach, feature components are extracted from different sub-modules of the images and concatenated to make a single vector to represent a face region. By doing this we are able to extract features that are more representative of the local features of the face. When projected onto an arbitrary line, samples from well formed clusters could produce a confused mixture of samples from all the classes leading to poor recognition. Discriminant analysis aims to find an optimal line orientation for which the data classes are well separated. Experiments performed on various databases to evaluate the performance of the proposed face recognition technique have shown improvement in recognition accuracy, especially under varying illumination conditions and facial expressions. This shows that the integration of multiple subspaces, each representing a part of a higher order nonlinear function, could represent a pattern with variability. Research work is progressing to investigate the effectiveness of subspace projection methodology for building manifolds with other nonlinear functions and to identify the optimum nonlinear function from an object classification perspective.

A multi-resolution approach for spinal metastasis detection using deep Siamese neural networks

Purpose In order to analyze the text complexity of Chinese and foreign academic English writings, the artificial neural network (ANN) under deep learning (DL) is applied to the study of text complexity. Firstly, the research status and existing problems of text complexity are introduced based on DL. Secondly, based on Back Propagation Neural Network (BPNN) algorithm, analyzation is made on the text complexity of Chinese and foreign academic English writings. And the research establishes a BPNN syntactic complexity evaluation system. Thirdly, MATLAB2013b is used for simulation analysis of the model. The proposed model algorithm BPANN is compared with other classical algorithms, and the weight value of each index and the model training effect are further analyzed by statistical methods. Finally, L2 Syntactic Complexity Analyzer (L2SCA) is used to calculate the syntactic complexity of the two libraries, and Mann–Whitney U test is used to compare the syntactic complexity of Chinese English learners and native English speakers. The experimental results show that compared with the shallow neural network, the deep neural network algorithm has more hidden layers and richer features, and better performance of feature extraction. BPNN algorithm shows excellent performance in the training process, and the actual output value is very close to the expected value. Meantime, the error of sample test is analyzed, and it is found that the evaluation error of BPNN algorithm is less than 1.8%, of high accuracy. However, there are significant differences in grammatical complexity among students with different English writing proficiency. Some measurement methods cannot effectively reflect the types and characteristics of written language, or may have a negative relationship with writing quality. In addition, the research also finds that the measurement of syntactic complexity is more sensitive to the language ability of writing. Therefore, BPNN algorithm can effectively analyze the text complexity of academic English writing. The results of the research provide reference for improving the evaluation system of text complexity of academic paper writing. Design/methodology/approach In order to analyze the text complexity of Chinese and foreign academic English writings, the artificial neural network (ANN) under deep learning (DL) is applied to the study of text complexity. Firstly, the research status and existing problems of text complexity are introduced based on DL. Secondly, based on Back Propagation Neural Network (BPNN) algorithm, analyzation is made on the text complexity of Chinese and foreign academic English writings. And the research establishes a BPNN syntactic complexity evaluation system. Thirdly, MATLAB2013b is used for simulation analysis of the model. The proposed model algorithm BPANN is compared with other classical algorithms, and the weight value of each index and the model training effect are further analyzed by statistical methods. Finally, L2 Syntactic Complexity Analyzer (L2SCA) is used to calculate the syntactic complexity of the two libraries, and Mann–Whitney U test is used to compare the syntactic complexity of Chinese English learners and native English speakers. The experimental results show that compared with the shallow neural network, the deep neural network algorithm has more hidden layers and richer features, and better performance of feature extraction. BPNN algorithm shows excellent performance in the training process, and the actual output value is very close to the expected value. Meantime, the error of sample test is analyzed, and it is found that the evaluation error of BPNN algorithm is less than 1.8%, of high accuracy. However, there are significant differences in grammatical complexity among students with different English writing proficiency. Some measurement methods cannot effectively reflect the types and characteristics of written language, or may have a negative relationship with writing quality. In addition, the research also finds that the measurement of syntactic complexity is more sensitive to the language ability of writing. Therefore, BPNN algorithm can effectively analyze the text complexity of academic English writing. The results of the research provide reference for improving the evaluation system of text complexity of academic paper writing. Findings In order to analyze the text complexity of Chinese and foreign academic English writings, the artificial neural network (ANN) under deep learning (DL) is applied to the study of text complexity. Firstly, the research status and existing problems of text complexity are introduced based on DL. Secondly, based on Back Propagation Neural Network (BPNN) algorithm, analyzation is made on the text complexity of Chinese and foreign academic English writings. And the research establishes a BPNN syntactic complexity evaluation system. Thirdly, MATLAB2013b is used for simulation analysis of the model. The proposed model algorithm BPANN is compared with other classical algorithms, and the weight value of each index and the model training effect are further analyzed by statistical methods. Finally, L2 Syntactic Complexity Analyzer (L2SCA) is used to calculate the syntactic complexity of the two libraries, and Mann–Whitney U test is used to compare the syntactic complexity of Chinese English learners and native English speakers. The experimental results show that compared with the shallow neural network, the deep neural network algorithm has more hidden layers and richer features, and better performance of feature extraction. BPNN algorithm shows excellent performance in the training process, and the actual output value is very close to the expected value. Meantime, the error of sample test is analyzed, and it is found that the evaluation error of BPNN algorithm is less than 1.8%, of high accuracy. However, there are significant differences in grammatical complexity among students with different English writing proficiency. Some measurement methods cannot effectively reflect the types and characteristics of written language, or may have a negative relationship with writing quality. In addition, the research also finds that the measurement of syntactic complexity is more sensitive to the language ability of writing. Therefore, BPNN algorithm can effectively analyze the text complexity of academic English writing. The results of the research provide reference for improving the evaluation system of text complexity of academic paper writing. Originality/value In order to analyze the text complexity of Chinese and foreign academic English writings, the artificial neural network (ANN) under deep learning (DL) is applied to the study of text complexity. Firstly, the research status and existing problems of text complexity are introduced based on DL. Secondly, based on Back Propagation Neural Network (BPNN) algorithm, analyzation is made on the text complexity of Chinese and foreign academic English writings. And the research establishes a BPNN syntactic complexity evaluation system. Thirdly, MATLAB2013b is used for simulation analysis of the model. The proposed model algorithm BPANN is compared with other classical algorithms, and the weight value of each index and the model training effect are further analyzed by statistical methods. Finally, L2 Syntactic Complexity Analyzer (L2SCA) is used to calculate the syntactic complexity of the two libraries, and Mann–Whitney U test is used to compare the syntactic complexity of Chinese English learners and native English speakers. The experimental results show that compared with the shallow neural network, the deep neural network algorithm has more hidden layers and richer features, and better performance of feature extraction. BPNN algorithm shows excellent performance in the training process, and the actual output value is very close to the expected value. Meantime, the error of sample test is analyzed, and it is found that the evaluation error of BPNN algorithm is less than 1.8%, of high accuracy. However, there are significant differences in grammatical complexity among students with different English writing proficiency. Some measurement methods cannot effectively reflect the types and characteristics of written language, or may have a negative relationship with writing quality. In addition, the research also finds that the measurement of syntactic complexity is more sensitive to the language ability of writing. Therefore, BPNN algorithm can effectively analyze the text complexity of academic English writing. The results of the research provide reference for improving the evaluation system of text complexity of academic paper writing.

Deep learning methods are widely used in the field of rolling bearing fault diagnosis and produce good results when faced with datasets with roughly equal numbers of normal and faulty samples. However, real-world data often has a serious imbalance, with the number of fault samples being significantly less than the number of normal samples. This dataset imbalance challenges the performance of traditional deep learning methods. To address this problem, this paper proposes an efficient imbalanced data rolling bearing fault diagnosis method. The method consists of two parts: a deep learning network based on a multi-scale self-attention mechanism and a novel loss function. In terms of the deep learning network, firstly, the one-dimensional vibration signal is converted into a two-dimensional image through the Gramian angular field. This conversion maximises the inherent feature extraction capability of the network. Subsequently, the multi-scale learning capability of the network is enhanced by implementing different expansion rates for the head of the multi-scale self-attention mechanism. This nuanced approach allows the network to capture the underlying information more efficiently. Finally, the inclusion of Ghost bottlenecks and feature pyramid networks (FPNs) helps to optimise network efficiency and improve generalisation performance. A novel loss function is also proposed to make the method more suitable for imbalanced data. During the training process, the classification of samples in each class is analysed using the recall metric of imbalanced classification and the real-time recall is used as a weight to weaken the dominance of the majority class. This weighting ensures the adaptability of the method to imbalanced datasets. The proposed method is evaluated using rolling bearing datasets from Case Western Reserve University, USA, and Southeast University, China. Comparison results with other state-of-the-art deep learning methods show that the proposed method has a robust performance when dealing with imbalanced data.

Deep learning has become widely used in image analysis. Transfer learning can make use of information from other datasets for the analysis of the chosen dataset. When there is a small number of images at hand, transfer learning using pre-trained models with coefficients already estimated from other datasets is recommended. This is in contrast to deep learning, where most model parameters are re-estimated. Deep transfer learning uses pre-trained models with fixed weight parameters in the lower layers; thus, deep learning can be viewed as a two-stage approach: (1) feature extraction from a lower neural network layer and (2) estimation of a neural network using the extracted features as inputs. Since deep transfer learning is feature extraction, we can extend the two-stage approach to a more general two-stage framework: (1) feature extraction using multiple methods and (2) machine learning methods taking extracted features as inputs. We evaluate the performance of methods with different Stage 1 and Stage 2 approaches in predicting the phenotype leaf numbers based on a multi-view plant imaging dataset. This paper contains an evaluation of different two-stage machine learning methods for multi-view image data in plant image phenotyping.