Final Softmax Layer Research Articles

Lightweight network based features fusion for steel rolling ambient sound classification

With the intelligent development of industrial production, sound monitoring technology has been widely used to monitor the operation status of mechanical facilities, and this progress has gradually become a research hotspot in the steel manufacturing industry. However, the complex composition of sound sources and the high computational requirements of most models limit their applicability to industrial scenarios. This work presents a targeted approach for remote monitoring of rolling sound. They are using visual features of audio signals to design classification models on the self-harvested dataset for more efficient adaptation to complex production sites. An in-depth analysis of the actual sound reveals that it is characterized by high similarity, complexity, and partial synchronization. To optimize interclass and intraclass data engineering few-sample, multi-feature fusion, and data augmentation methods were combined to characterize the details fully. In addition, by leveraging the deep small convolutions of VGGNet and the randomness of stochastic pooling, local features are effectively extracted. Finally, global average pooling with final softmax layer stacking is used to classify the rolling signals to reduce the number of parameters, avoid overfitting, and perform a global analysis of the features. Experimental results on the rolling sound dataset show that the method proposed achieves an accuracy of 91.26%, with 92.66%, and 95.88% on the ESC10 and MIMII datasets, respectively. These results confirm that the method can be widely applied to sound classification for multi-category rolling processes, showing good performance and scalability.

Research on fault diagnosis of rolling bearing based on improved convolutional neural network with sparrow search algorithm.

Traditional approaches to the intelligent fault diagnosis of rolling bearings have predominantly relied on manual expertise for feature extraction, a practice that compromises robustness. In addition, the existing convolutional neural network (CNN) is characterized by an overabundance of parameters and a substantial requirement for training samples. To address these limitations, this study introduces a novel fault diagnosis algorithm for rolling bearings, integrating a one-dimensional convolutional neural network (1DCNN) with a support vector machine (SVM) to form an enhanced 1DCNN-SVM model. This model is further refined using the sparrow search algorithm (SSA) for the optimal adjustment of the parameters of 1DCNN-SVM. Specifically, by substituting the CNN's final softmax layer with an SVM, the model becomes better suited for processing limited data volumes. In addition, the incorporation of batch normalization and dropout layers within the CNN framework significantly augments its fault classification accuracy for rolling bearings, concurrently mitigating the risk of overfitting. The SSA is subsequently applied to refine three principal hyper-parameters: batch size, initial learning rate, and the L2 regularization coefficient, thereby overcoming the challenges associated with manually adjusting parameters, such as extended processing times and unpredictable outcomes. Empirical tests on Case Western Reserve University (CWRU) datasets revealed the model's superior performance, with the SSA-optimized 1DCNN-SVM showcasing diagnostic accuracies over 98%, marked improvements over conventional models, and a significant reduction in processing times. This method not only marks a significant advancement in intelligent fault diagnosis for rolling bearings but also demonstrates the potential of integrating machine learning for more precise and efficient diagnostics. The SSA-1DCNN-SVM model, optimized for accuracy and minimal data use, sets a new standard in fault diagnosis, relevant for machinery health monitoring and maintenance strategies across various industries.

CNN based image resizing forensics for double compressed JPEG images

This paper investigates a novel CNN-based architecture for image resizing forensics in the presence of Double-JPEG compression. Two sub-problems are addressed as part of this paper: first, the detection of resizing in DJPEG images, and second, determining the factor used to resize the image before the second JPEG compression. The image resizing technique used in this paper is image scaling. The proposed network is based on utilizing specially chosen high pass filters that constitute the pre-processing layer, followed by four convolutional layers, two fully connected layers, and a final softmax layer. The high-pass filters preserve the resizing clues in the form of residuals applied at the pre-processing layer, followed by utilizing the residuals as inputs to a CNN to detect the presence of resizing and also for estimating the resizing factor. The proposed network is fully end-to-end and does not rely on any hand-crafting. Experimental validation is carried out utilizing images from the RAISE dataset. Comparative analysis with four state-of-the-art methods demonstrates the proposed CNN-based architecture’s superiority for resizing detection and resize factor determination sub-problems.

Abstract 11242: Spatial-Temporal Deep Learning for Automatic Detection of Echocardiographic Predictors of Mortality in Patients Hospitalized With Covid-19 - Data From the PROVAR-COVID Study

Background: Cardiac involvement seems to impact the prognosis of COVID-19. Bedside echocardiography (echo) holds promise for the early outcome prediction, and artificial intelligence has shown to be an additional tool to overcome personnel limitations. We propose a spatial-temporal deep learning-based approach for automatic prediction of mortality of inpatients with COVID-19 with echo images. Methods: Patients admitted to two hospitals in Brazil in 90 days with confirmed moderate and severe COVID-19 underwent clinical and laboratory evaluation, and focused bedside echo (GE Vivid IQ), following admission, with remote interpretation. Independent echo predictors of all-cause mortality were assessed, after adjustment for clinical variables. Our image dataset consists of 752 videos (Mpeg) acquired in three different probe positions: apical 4-chamber and parasternal long and short axis. We predicted the patients’ outcome (discharge or death) in a 10-fold cross-validation procedure manner, where we applied a 2-stream deep neural network composed of 2 identical ResNet-18 Convolutional Neural Networks (CNN) - pre-trained and developed for Rheumatic Heart Disease - followed by attention units to extract relevant spatial and temporal features from RGB and Optical Flow frames. A final softmax layer is used for classification. Results: A total of 163 patients were enrolled, the mean age was 64±16 years, 107 (66%) were admitted to intensive care and in-hospital mortality was 34% (N=56). Independent predictors of mortality were age≥63 years (OR=5.53, 95%CI 1.52-20.17), LVEF<64% (OR=7.37, 95%CI 2.10-25.94), and TAPSE<18.5 mm (OR=9.43, 95% CI 2.57-35.03), C-statistic=0.83. During training, utilizing a balanced set of 106 patients, we rescaled videos to 224x224 pixels and used temporal jittering to select a final clip with 32 frames per video in random batches of size 6. Our proposed method achieved an exam-wise accuracy of 62.22% (95%CI 52.40-72,04), sensitivity of 67.52% (95%CI 51.62-83.42), and specificity of 60.45% (95%CI 45.15-75.75) for mortality. Conclusion: Automatic detection of high-risk echo findings in COVID-19 inpatients at the bedside seems feasible and, with more research, can improve mortality prediction at the point of care.

Stacked auto-encoders based visual features for speech/music classification

With the rapid rise of online available content, multimedia signal processing has become an important area of research. The output of the speech/music classifier (SMC) is further used for both speech and music signal processing. Most of the works in the past two decades are focused on conventional classifiers mostly using shallow structures. This work evaluates the performance of stacked auto-encoders based deep neural network for designing a speech/music classifier on S&S and GTZAN dataset using visual features. The hidden layers of the neural network are initially trained in unsupervised manner using auto-encoders and are stacked with the final softmax layer. Different experiments were conducted on time-frequency features derived from Spectrogram and Chromagram. Performances of the combination of stacked auto-encoder and softmax classifier was further compared with traditional classifiers and different deep learning techniques. Best classification accuracy of 93.05% and 94.73% is observed for fused features for S&S and GTZAN datasets respectively.

Attention-Aware Pseudo-3-D Convolutional Neural Network for Hyperspectral Image Classification

Convolutional neural networks (CNNs) have been applied for hyperspectral image classification recently. Among this class of deep models, 3-D CNN has been shown to be more effective by learning discriminative features from abundant spectral signatures and spatial contexts in hyperspectral imagery (HSI). However, by simply imposing 3-D CNN to HSI, a large amount of initial information might be lost in this CNN pipeline. The proposed attention-aware pseudo-3-D (AP3D) convolutional network for HSI classification is motivated by two observations. First, each dimension of the 3-D HSI is not equally important, different attention should be paid to different dimensions of the initial HSI image, especially in the first convolution operation. Second, intermediate representations of the 3-D input image at different stages in the 3-D CNN pipeline represent different levels of features and should not be neglected and abandoned. Instead, a 2-D matrix of scores for each feature map should be fed to the final softmax layer. Quantitative and qualitative results demonstrate that the proposed AP3D model outperforms the state-of-the-art HSI classification methods in agricultural and rural/urban data sets: Indian Pines, Pavia University, and Salinas Scene.

Optimized convolutional neural network model for plant species identification from leaf images using computer vision

In recent works of computer science, especially in the fields of image processing and pattern recognition techniques with machine learning, considerable focus is given to plant taxonomy which enhances the abilities of people to recognize plant species. This paper presents a method that analyzes color images of leaves using a type of Convolutional Neural Network to recognize plant species. The proposed Neural Network consists of four convolutional layers followed by two Fully-Connected layers and a final soft-max layer to offer a feature representation for different plant species. Four max-pooling layers are performed over a 2 × 2 pixel window with stride 2. Results on five plant datasets viz. Leaf snap (52 plant species), UCI leaf (40 plant species), PlantVillage (38 plant species), Flavia (32 plant species) and Swedish (15 plant species) are tabulated that demonstrate the remarkable performance of the proposed deep neural network when compared to the state of art methods.

Deep neural network for semi-automatic classification of term and preterm uterine recordings.

Pregnancy is a complex process, and the prediction of premature birth is uncertain. Many researchers are exploring non-invasive approaches to enhance its predictability. Currently, the ElectroHysteroGram (EHG) and Tocography (TOCO) signal are a real-time and non-invasive technology which can be employed to predict preterm birth. For this purpose, sparse autoencoder (SAE) based deep neural network (SAE-based DNN) is developed. The deep neural network has three layers including a stacked sparse autoencoder (SSAE) network with two hidden layers and one final softmax layer. To this end, the bursts of all 26 recordings of the publicly available TPEHGT DS database corresponding to uterine contraction intervals and non-contraction intervals (dummy intervals) were manually segmented. 20 features were extracted by two feature extraction algorithms including sample entropy and wavelet entropy. Afterwards, the SSAE network is adopted to learn high-level features from raw features by unsupervised learning. The softmax layer is added at the top of the SSAE network for classification. In order to verify the effectiveness of the proposed method, this study used 10-fold cross-validation and four indicators to evaluate classification performance. Experimental research results display that the performance of deep neural network can achieve Sensitivity of 98.2%, Specificity of 97.74%, and Accuracy of 97.9% in the publicly TPEHGT DS database. The performance of deep neural network outperforms the comparison models including deep belief networks (DBN) and hierarchical extreme learning machine (H-ELM). Finally, experimental research results reveal that the proposed method could be valid applied to semi-automatic identification of term and preterm uterine recordings.

A Novel Counterfeit Feature Extraction Technique for Exposing Face-Swap Images Based on Deep Learning and Error Level Analysis.

The quality and efficiency of generating face-swap images have been markedly strengthened by deep learning. For instance, the face-swap manipulations by DeepFake are so real that it is tricky to distinguish authenticity through automatic or manual detection. To augment the efficiency of distinguishing face-swap images generated by DeepFake from real facial ones, a novel counterfeit feature extraction technique was developed based on deep learning and error level analysis (ELA). It is related to entropy and information theory such as cross-entropy loss function in the final softmax layer. The DeepFake algorithm is only able to generate limited resolutions. Therefore, this algorithm results in two different image compression ratios between the fake face area as the foreground and the original area as the background, which would leave distinctive counterfeit traces. Through the ELA method, we can detect whether there are different image compression ratios. Convolution neural network (CNN), one of the representative technologies of deep learning, can extract the counterfeit feature and detect whether images are fake. Experiments show that the training efficiency of the CNN model can be significantly improved by the ELA method. In addition, the proposed technique can accurately extract the counterfeit feature, and therefore achieves outperformance in simplicity and efficiency compared with direct detection methods. Specifically, without loss of accuracy, the amount of computation can be significantly reduced (where the required floating-point computing power is reduced by more than 90%).

Plant identification based on very deep convolutional neural networks

Plant identification is a critical step in protecting plant diversity. However, many existing identification systems prohibitively rely on hand-crafted features for plant species identification. In this paper, a deep learning method is employed to extract discriminative features from plant images along with a linear SVM for plant identification. To offer a self-learning feature representation for different plant organs, we choose a very deep convolutional neural networks (CNNs), which consists of sixteen convolutional layers followed by three Fully-Connected (FC) layers and a final soft-max layer. Five max-pooling layers are performed over a 2?2 pixel window with stride 2. Extensive experiments on several plant datasets demonstrate the remarkable performance of the very deep neural network compared to the hand-crafted features.