Electronic noses (e-noses) have proven effective in detecting lung cancer by analyzing Volatile Organic Compounds (VOCs) in breath samples, with Multivariate Time Series Classification (MTSC) as the primary task. However, challenges remain in effectively capturing spatiotemporal information for MTSC. To address this, a novel method called SIR-3DCNN for MTSC in lung cancer detection is introduced. A pivotal aspect of SIR-3DCNN is the Sensor Array Optimization (SAO), an algorithm based on Linear Discriminant Analysis (LDA) that can decrease the number of sensors from 22 to 8 while increasing accuracy by 2.35 percentage points in our study. Furthermore, SIR-3DCNN incorporates an advanced technique for representing spatiotemporal information, converting optimized Multivariate Time Series (MTS) into Maximum Trajectory Matrix Images (MTMIs) and arranging them to maximize the Sum of Inter-Frame Mutual Information (SIFMI). Additionally, we have developed C3D-Light, a lightweight yet effective 3D Convolutional Neural Network (3DCNN) that demonstrates superior performance compared to other models. Comparative analyses with state-of-the-art methodologies reveal that SIR-3DCNN consistently outperforms existing methods, achieving perfect sensitivity (100%), the highest specificity (80%), accuracy (92.94%), AUC (94.85%), precision (90.26%), and F1-score (94.86%) among all compared methods. This advancement holds significant promise for lung cancer detection using electronic noses. The source code is available at https://github.com/cqu-3dteam/sir-3dcnn.

The electronic nose (e-nose) is susceptible to sensor drift and instrumental variation, which may result in distribution discrepancy in data collected, hence leading to classification performance degradation. It is necessary to apply domain adaptation to solve this problem. A nonlinear subspace projection approach named Feature Entropy Domain Adaption (FEDA) is proposed for domain adaptation for e-nose data classification. One important aspect of FEDA is that adversarial training is introduced to minimize the distribution discrepancy between source and target domains. No projection matrix and parameter fine-tuning are needed anymore in comparison with the popular linear subspace projection approaches. In addition, feature norm and conditional entropy are introduced into adversarial training in FEDA to reduce the decision boundary uncertainty and the overlap between classes, respectively. Experimental results show that FEDA can deal with the distribution discrepancy of e-nose effectively, and can achieve satisfactory classification accuracy on various datasets. Source code can be found at https://github.com/threedteam/DA_FEDA.

Due to sensor drift, instrumental variation, or change of measurement object, the distributions of the datasets (domains) acquired by the sensors are often different. A domain adaptive framework called MCSP-SSS is proposed to reduce the distribution discrepancy across domains for high-accuracy sensor data classification. The framework consists of two key parts: Multi-Constraint Subspace Projection (MCSP) and Source Sample Selection (SSS). MCSP is a subspace-projection-based approach, which introduces four constraints to get an optimized projection matrix: Principal Component Analysis (PCA) is used to reduce the redundancy of features; Mean Distribution Discrepancy (MDD) is applied to minimize the difference between source and target data; Hilbert-Schmidt Independence Criterion (HSIC) is adopted to maximizing the dependence between features and labels; A so-called weighted-within-class scatter matrix is introduced to minimize the within-class variance to avoid samples with different labels to overlap in the subspace. SSS is designed to remove the outliers in projected source samples so as to reduce distribution discrepancy between the source and target samples projected by MCSP. Experimental results show that our framework can achieve the best accuracies in all classification tasks in comparison with other state-of-the-art approaches. The accuracy of MCSP-SSS is 3.57 percentage points (pps) higher on average than that of the second best approach for single-source domain adaptation, and 7.00 pps for multi-source domain adaptation. Source code is available at https://github.com/threedteam/tsc_subspace_projection .

Sensor data classification, a typical Time Series Classification (TSC) problem, spans many applications of sensors in the real world. How to improve classification performance is a challenging problem in these applications. A method based on Phase-Space Reconstruction (PSR), named MTMI-DCNN, is proposed for TSC. The basic idea of this method is to convert the time series to 2D images with the largest possible resolution by PSR, and then use Deep Convolutional Neural Network (DCNN) to classify these images. Three sequential stages are contained in this method: parameter optimization stage, time-series-imaging stage, and classification stage. In the parameter optimization stage, an optimization method is presented to optimize the reconstruction parameters just by maximizing the resolution of the image converted. In the time-series-imaging stage, the time series is embedded into a phase space using the optimal parameters, and the trajectory matrix of the time series is treated directly as an image for later classification. Consequently, no phase-space-trajectory projection is needed and the information loss caused by the projection is avoided. In the classification stage, the images converted from time series are classified by DCNN. Experimental results show that the proposed MTMI-DCNN outperforms other state-of-the-art methods on various datasets: the accuracies achieve 98.7%, 100.0%, and 82.6% on datasets UoB, Coffee, and CHMO, respectively; the Number of Wins and Average Error Rate achieve 19 and 1.41% respectively on 60 UTS datasets from the UCR archive. Hence, it can serve as an additional valuable method for TSC. Source code is available at https://github.com/threedteam/MTMI-DCNN/ .

The Depth-Image-Based Rendering (DIBR) algorithm is pivotal in the advancement of Virtual Reality (VR) and Augmented Reality (AR) technologies due to its capacity to generate virtual views from arbitrary perspectives. Nonetheless, the generation process is often marred by the occurrence of holes due to sharp depth transitions, significantly degrading the quality of the synthesized view. To mitigate this issue, this study introduces a layered hole-filling method to enhance the quality of virtual views. The effectiveness of our proposed method is ensured through three key techniques: Firstly, a depth-aware decomposition technique is employed to precisely segregate foreground objects from the background within a reference view. This is achieved by leveraging both the reference image and its corresponding depth map, facilitating accurate instance-level separation of foreground objects. Secondly, a Generative Adversarial Network (GAN)-enhanced background reconstruction technique is proposed to generate hole-free target views devoid of foreground objects. Lastly, the integration of Masked 3D Image Warping (M3DIW) and Layered Mergence (LM) algorithms facilitates filling holes with foreground or background textures in a layered manner. Comprehensive experimental results demonstrate the superiority of our proposed method compared to state-of-the-art methods. Notably, our method demonstrates an improvement of 7.5% in mean average Peak Signal-to-Noise Ratio (PSNR) and 1.8% in mean average Structural Similarity Index Measure (SSIM) compared to existing techniques. Additionally, it impressively lowers mean average Learned Perceptual Image Patch Similarity (LPIPS) by 28.8% and significantly reduces mean average Fréchet Inception Distance (FID) by 28.9% for all sequences tested. These results affirm the effectiveness of our approach in enhancing the quality of virtual view synthesis within DIBR applications. Source code is available at https://github.com/threedteam/dibr.