Deep Learning Frameworks Evaluation for Image Classification on Resource Constrained Device

PurposeThe purpose of this study is to explore the potential of trainable activation functions to enhance the performance of deep neural networks, specifically ResNet architectures, in the task of image classification. By introducing activation functions that adapt during training, the authors aim to determine whether such flexibility can lead to improved learning outcomes and generalization capabilities compared to static activation functions like ReLU. This research seeks to provide insights into how dynamic nonlinearities might influence deep learning models' efficiency and accuracy in handling complex image data sets.Design/methodology/approachThis research integrates three novel trainable activation functions – CosLU, DELU and ReLUN – into various ResNet-n architectures, where “n” denotes the number of convolutional layers. Using CIFAR-10 and CIFAR-100 data sets, the authors conducted a comparative study to assess the impact of these functions on image classification accuracy. The approach included modifying the traditional ResNet models by replacing their static activation functions with the trainable variants, allowing for dynamic adaptation during training. The performance was evaluated based on accuracy metrics and loss profiles across different network depths.FindingsThe findings indicate that trainable activation functions, particularly CosLU, can significantly enhance the performance of deep learning models, outperforming the traditional ReLU in deeper network configurations on the CIFAR-10 data set. CosLU showed the highest improvement in accuracy, whereas DELU and ReLUN offered varying levels of performance enhancements. These functions also demonstrated potential in reducing overfitting and improving model generalization across more complex data sets like CIFAR-100, suggesting that the adaptability of activation functions plays a crucial role in the training dynamics of deep neural networks.Originality/valueThis study contributes to the field of deep learning by introducing and evaluating the impact of three novel trainable activation functions within widely used ResNet architectures. Unlike previous works that primarily focused on static activation functions, this research demonstrates that incorporating trainable nonlinearities can lead to significant improvements in model performance and adaptability. The introduction of CosLU, DELU and ReLUN provides a new pathway for enhancing the flexibility and efficiency of neural networks, potentially setting a new standard for future deep learning applications in image classification and beyond.

Retinal disorders, including diabetic retinopathy and macular degeneration due to aging, can lead to preventable blindness in diabetics. Vision loss caused by diseases that affect the retinal fundus cannot be reversed if not diagnosed and treated on time. This paper employs deep-learned feature extraction with ensemble learning models to improve the multi-disease classification of fundus images. This research presents a novel approach to the multi-classification of fundus images, utilizing deep-learned feature extraction techniques and ensemble learning to diagnose retinal disorders and diagnosing eye illnesses involving feature extraction, classification, and preprocessing of fundus images. The study involves analysis of deep learning and implementation of image processing. The ensemble learning classifiers have used retinal photos to increase the classification accuracy. The results demonstrate improved accuracy in diagnosing retinal disorders using DL feature extraction and ensemble learning models. The study achieved an overall accuracy of 87.2%, which is a significant improvement over the previous study. The deep learning models utilized in the study, including NASNetMobile, InceptionResNetV4, VGG16, and Xception, were effective in extracting relevant features from the Fundus images. The average F1-score for Extra Tree was 99%, while for Histogram Gradient Boosting and Random Forest, it was 98.8% and 98.4%, respectively. The results show that all three algorithms are suitable for the classification task. The combination of DenseNet feature extraction technique and RF, ET, and HG classifiers outperforms other techniques and classifiers. This indicates that using DenseNet for feature extraction can effectively enhance the performance of classifiers in the task of image classification.

With the recent development of deep learning techniques, deep learning methods are widely used in image classification tasks, especially for those based on convolutional neural networks (CNN). In this paper, a general overview on the image classification tasks will be presented. Besides, the differences and contributions to essential progress in the image classification tasks of the deep learning models including LeNet, AlexNet, Inception, VggNet and ResNet are introduced. This paper will also explain in detail, how different units in these CNN models, other than the convolutional layer, including pooling, activation, and dropout functionalize to support better results for these models. These results offer a guideline for deeply understanding the utility of CNN units.

In image classification tasks, commonly used deep learning network architectures include Convolutional Neural Network (CNN) and Vision Transformer (ViT). Still, both are relatively mature, while Graph Neural Network (GNN) network architecture has been attempted less in image classification. In addition, the architecture and design of GNN are still rapidly developing and have enormous potential. This study aims to create an image classification method based on superpixels and graph neural networks. This project attempted three different approaches to preserve superpixel node features with varying numbers of features, designed two training network architectures with different complexities and conducted six sets of tests on image classification tasks on the fashion-mnist dataset. As a result, the use of superpixel and GNN methods achieved good accuracy in image classification tasks, demonstrating the potential of this method in image recognition tasks. It was also found that adding the boundary coordinates of the region to the training in the selection of superpixel node features can improve the accuracy of the final image classification task.

Deep neural networks have achieved remarkable results in the field of polarimetric synthetic aperture radar (PolSAR) image classification. However, PolSAR is affected by speckle imaging, resulting in PolSAR images usually containing a large amount of speckle noise, which usually leads to the poor spatial consistency of classification results and insufficient classification accuracy. Semantic segmentation methods based on deep learning can realize the task of image segmentation and classification at the same time, producing fine-grained and smooth classification maps. However, these approaches require enormous labeled data sets, which are laborious and time-consuming. Due to these issues, a new multi-modal contrastive fully convolutional network, named MCFCN, is proposed for PolSAR image classification in this paper, which combines multi-modal features of the same pixel as inputs to the model based on a fully convolutional network and accomplishes the classification task using only a small amount of labeled data through contrastive learning. In addition, to describe the PolSAR terrain targets more comprehensively and enhance the robustness of the classifier, a pixel overlapping classification strategy is proposed, which can not only improve the classification accuracy effectively but also enhance the stability of the method. The experiments demonstrate that compared with existing classification methods, the classification results of the proposed method for three real PolSAR datasets have higher classification accuracy.

Snake Species Identification is a challenge as erroneous snake identification from the perceptible traits is a prime reason of death because of snake bites. The main objective of the proposed system is to be able to identify snake species from their visual traits in order to provide suitable treatment, thus preventing subsequent deaths. The proposed system involves techniques based on Image Processing, Convolution Neural Networks and Deep Learning to achieve the mentioned purpose. CNN has been highly used in automatic image classification system. In most cases, extracting features and utilizing them for classification. Deep learning successfully achieves recognition of objects in images as it is implemented using artificial neural networks. Image classification tasks have seen a rise with the introduction of deep learning techniques. So far, no automated method for classification has been suggested to categorize snakes. The system that would be developed will be useful to recognize snake species correctly and thus take necessary action.

Convolutional neural networks (CNNs), the most important deep learning networks for computer vision, have undergone a series of developments and improvements for image-related tasks such as object recognition, image classification, semantic segmentation, etc. However, in the field of natural language processing (NLP), the novel attention-based network Transformer had a profound impact on machine translation, which subsequently led to a boom in attention-based models for computer vision. State-of-the-art models with attention have already shown good performance for computer vision tasks due to the sophisticated design of network architectures and advanced computational efficiency techniques. For example, selfattention learns relationships between segments or words in different positions compared to the current performance of convolutional neural networks. Inspired by Vision Transformer (ViT), we propose a simple novel transformer architecture model, called Flexible Transformer, which inherits the properties of attention-based architectures and is flexible for inputs of arbitrary size. Besides self-attention, the inputs in ViT are not pre-processed, such as resizing or cropping, but kept intact without altering them, which could lead to distortion or loss of information. In this paper, we want to present a novel and simple architecture that meets these requirements. Compared to state-of-arts, our model processes inputs with arbitrary image sizes without any pre-processing and pretraining costs. Also, the results of the experiments show that the model can potentially provide good results with high accuracy despite limited resources. Even though the results of Flexible Transformer are not as accurate as those of Vision Transformer, they show the potential of a model with high performance in image classification tasks with variable size images. The significance of the research opens up possibilities for dealing with primitive images in deep learning tasks. Based on the original inputs, reliable results with good accuracy could be obtained if the proposed model is optimized and further trained on large datasets.

To enhance the performance of image classification and speech recognition, the optimizer is considered an important factor for achieving high accuracy. The state-of-the-art optimizer can perform to serve in applications that may not require very high accuracy, yet the demand for high-precision image classification and speech recognition is increasing. This study implements an adaptive method for applying the particle filter technique with a gradient descent optimizer to improve model learning performance. Using a pretrained model helps reduce the computational time to deploy an image classification model and uses a simple deep convolutional neural network for speech recognition. The applied method results in a higher speech recognition accuracy score—89.693% for the test dataset—than the conventional method, which reaches 89.325%. The applied method also performs well on the image classification task, reaching an accuracy of 89.860% on the test dataset, better than the conventional method, which has an accuracy of 89.644%. Despite a slight difference in accuracy, the applied optimizer performs well in this dataset overall.

Fully connected multi layer neural networks such as Deep Boltzmann Machines (DBM) performs better than fully connected single layer neural networks in image classification tasks and has a smaller number of hidden layer neurons than Extreme Learning Machine (ELM) based fully connected multi layer neural networks such as Multi Layer ELM (MLELM) and Hierarchical ELM (H-ELM) However, ML-ELM and H-ELM has a smaller training time than DBM. This paper introduces a fully connected multi layer neural network referred to as Multi Layer Multi Objective Extreme Learning Machine (MLMO-ELM) which uses a multi objective formulation to pass the label and non-linear information in order to learn a network model which has a similar number of hidden layer parameters as DBM and smaller training time than DBM. The experimental results show that MLMO-ELM outperforms DBM, ML-ELM and H-ELM on OCR and NORB datasets.

Objective.Major depressive disorder (MDD) is a widespread mental disorder that affects health. Many methods combining electroencephalography (EEG) with machine learning or deep learning have been proposed to objectively distinguish between MDD and healthy individuals. However, most current methods detect depression based on multichannel EEG signals, which constrains its application in daily life. The context in which EEG is obtained can vary in terms of study designs and EEG equipment settings, and the available depression EEG data is limited, which could also potentially lessen the efficacy of the model in differentiating between MDD and healthy subjects. To solve the above challenges, a depression detection model leveraging transfer learning with the single-channel EEG is advanced.Approach.We utilized a pretrained ResNet152V2 network to which a flattening layer and dense layer were appended. The method of feature extraction was applied, meaning that all layers within ResNet152V2 were frozen and only the parameters of the newly added layers were adjustable during training. Given the superiority of deep neural networks in image processing, the temporal sequences of EEG signals are first converted into images, transforming the problem of EEG signal categorization into an image classification task. Subsequently, a cross-subject experimental strategy was adopted for model training and performance evaluation.Main results.The model was capable of precisely (approaching 100% accuracy) identifying depression in other individuals by employing single-channel EEG samples obtained from a limited number of subjects. Furthermore, the model exhibited superior performance across four publicly available depression EEG datasets, thereby demonstrating good adaptability in response to variations in EEG caused by the context.Significance.This research not only highlights the impressive potential of deep transfer learning techniques in EEG signal analysis but also paves the way for innovative technical approaches to facilitate early diagnosis of associated mental disorders in the future.

Currently, deep learning (DL) algorithms perform best in image classification and object detection tasks. Consequently, they are frequently used to address most problems involving computer vision. In this sense, the pervasive presence of smartphones and IoT devices has created a need to make this artificial intelligence portable. Given that deep neural network (DNN) models consist of millions of parameters, emerging research efforts have focused on enabling offline DL execution on low-resource devices, such as within the TinyML paradigm. This study analyzes the state-of-the-art of DL embedded in smartphones to develop an app for children that can recognize 3D geometric shapes without needing an internet connection. Alongside a systematic literature review, we conduct experiments with several pre-trained and lightweight models, which were subsequently evaluated using parametric statistical tests. While DL on smartphones is an underexplored area, it is expected to evolve significantly. Among the classification models tested, DenseNet169 demonstrated the highest accuracy (81%), whereas the MobileNet variants were faster and closer to real-time performance (30 FPS). In detection tasks, the EfficientDet-Lite and YOLOv8 models were evaluated, with EfficientDet-Lite being less accurate but faster (50 ms) compared to YOLOv8 (4 seconds). Although the field of DL on smartphones still requires further development, current lightweight models and frameworks offer significant opportunities for practical application.

This paper presents the ‘hyper-sinh’, a variation of the m-arcsinh activation function suit-able for Deep Learning (DL)-based algorithms for supervised learning, including Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), such as the Long Short-Term Memory (LSTM). hyper-sinh, developed in the open-source Python libraries TensorFlow and Keras, is thus described and validated as an accurate and reliable activation function for shallow and deep neural networks. Improvements in accuracy and reliability in image and text classification tasks on six (N=6) medium-to-large open-source benchmark datasets are discussed. Experimental results demonstrate that the overall competitive classification performance of the novel hyper-sinh function on shallow and deep neural networks yielded the highest performance. Furthermore, this activation is evaluated against other gold standard activation functions, demonstrating its overall competitive accuracy and reliability for both image and text classification tasks.

Deep neural networks achieve state-of-the-art performance for image classification and other tasks but are easily fooled by forgeries which slightly modify a legitimate image in a specific direction and are visually indistinguishable from the original. This presents a security risk for applications such as driverless transportation systems. We formulate detection of such forgeries as a watermark detection problem and derive locally optimal statistical tests for identifying them. Motivated by this optimal structure, we present a procedure for learning a forgery detector from a training set. The reliability of our forgery detector is assessed for several image classification tasks.

Although the existing traditional image classification methods have been widely applied in practical problems, there are some problems in the application process, such as unsatisfactory effects, low classification accuracy, and weak adaptive ability. This method separates image feature extraction and classification into two steps for classification operation. The deep learning model has a powerful learning ability, which integrates the feature extraction and classification process into a whole to complete the image classification test, which can effectively improve the image classification accuracy. However, this method has the following problems in the application process: first, it is impossible to effectively approximate the complex functions in the deep learning model. Second, the deep learning model comes with a low classifier with low accuracy. So, this paper introduces the idea of sparse representation into the architecture of the deep learning network and comprehensively utilizes the sparse representation of well multidimensional data linear decomposition ability and the deep structural advantages of multilayer nonlinear mapping to complete the complex function approximation in the deep learning model. And a sparse representation classification method based on the optimized kernel function is proposed to replace the classifier in the deep learning model, thereby improving the image classification effect. Therefore, this paper proposes an image classification algorithm based on the stacked sparse coding depth learning model-optimized kernel function nonnegative sparse representation. The experimental results show that the proposed method not only has a higher average accuracy than other mainstream methods but also can be good adapted to various image databases. Compared with other deep learning methods, it can better solve the problems of complex function approximation and poor classifier effect, thus further improving image classification accuracy.

Large-scale datasets are essential for the success of deep learning in image retrieval. However, manual assessment errors and semi-supervised annotation techniques can lead to label noise even in popular datasets. As previous works primarily studied annotation quality in image classification tasks, it is still unclear how label noise affects deep learning approaches to image retrieval. In this work, we show that image retrieval methods are less robust to label noise than image classification ones. Furthermore, we, for the first time, investigate different types of label noise specific to image retrieval tasks and study their effect on model performance.