• Parameter-free encoder block uses noisy transforms to improve generalization. • Fusing clean and distorted features enhances model representation capabilities and explainability. • Frequency-based attention leverages high-frequency data for boundary focus. • Extensive experiments on SOTA architectures validate the proposed approach.

With the rapid development of the Internet-of-Medical-Things (IoMT) in recent years, it has emerged as a promising solution to alleviate the workload of medical staff, particularly in the field of Medical Image Quality Assessment (MIQA). By deploying MIQA based on IoMT, it proves to be highly valuable in assisting the diagnosis and treatment of various types of medical images, such as fundus images, ultrasound images, and dermoscopic images. However, traditional MIQA models necessitate a substantial number of labeled medical images to be effective, which poses a challenge in acquiring a sufficient training dataset. To address this issue, we present a label-free MIQA model developed through a zero-shot learning approach. This paper introduces a Semantics-Aware Contrastive Learning (SCL) model that can effectively generalise quality assessment to diverse medical image types. The proposed method integrates features extracted from zero-shot learning, the spatial domain, and the frequency domain. Zero-shot learning is achieved through a tailored Contrastive Language-Image Pre-training (CLIP) model. Natural Scene Statistics (NSS) and patch-based features are extracted in the spatial domain, while frequency features are hierarchically extracted from both local and global levels. All of this information is utilised to derive a final quality score for a medical image. To ensure a comprehensive evaluation, we not only utilise two existing datasets, EyeQ and LiverQ, but also create a dataset specifically for skin image quality assessment. As a result, our SCL method undergoes extensive evaluation using all three medical image quality datasets, demonstrating its superiority over advanced models.

Unsustainable fishing, driven by bycatch and discards, harms marine ecosystems. Addressing this, we propose a Coordinate-Aware Mask R-CNN (CAM-RCNN) method to enhance fish detection in commercial trawls. Leveraging CoordConv and Group Normalization, our approach improves generalization and stability. To tackle class imbalance, a compound Dice and cross-entropy loss is employed, and image data are enhanced through multi-scale retinex and color restoration. Evaluating on two fishing datasets, CAM-RCNN excels in accuracy and generalization, achieving the best Average Precision (AP) for instance mask and BBOX prediction in both source (39.7%, 40.2%) and target domains (24.4%, 24.2%). This method promotes sustainable fishing by selectively capturing desired fish, reducing harm to non-target species.

Biomedical image segmentation plays an important role in Diabetic Retinopathy (DR)-related biomarker detection. DR is an ocular disease that affects the retina in people with diabetes and could lead to visual impairment if management measures are not taken in a timely manner. In DR screening programs, the presence and severity of DR are identified and classified based on various microvascular lesions detected by qualified ophthalmic screeners. Such a detection process is time-consuming and error-prone, given the small size of the microvascular lesions and the volume of images, especially with the increasing prevalence of diabetes. Automated image processing using deep learning methods is recognized as a promising approach to support diabetic retinopathy screening. In this paper, we propose a novel compound scaling encoder-decoder network architecture to improve the accuracy and running efficiency of microvascular lesion segmentation. In the encoder phase, we develop a lightweight encoder to speed up the training process, where the encoder network is scaled up in depth, width, and resolution dimensions. In the decoder phase, an attention mechanism is introduced to yield higher accuracy. Specifically, we employ Concurrent Spatial and Channel Squeeze and Channel Excitation (scSE) blocks to fully utilise both spatial and channel-wise information. Additionally, a compound loss function is incorporated with transfer learning to handle the problem of imbalanced data and further improve performance. To assess performance, our method is evaluated on two large-scale lesion segmentation datasets: DDR and FGADR datasets. Experimental results demonstrate the superiority of our method compared to other competent methods. Our codes are available at https://github.com/DeweiYi/CoSED-Net.

Humans and animals have the ability to continuously learn new information over their lifetime without losing previously acquired knowledge. However, artificial neural networks struggle with this due to new information conflicting with old knowledge, resulting in catastrophic forgetting. The complementary learning systems (CLS) theory (McClelland and McNaughton, 1995; Kumaran et al. 2016) suggests that the interplay between hippocampus and neocortex systems enables long-term and efficient learning in the mammalian brain, with memory replay facilitating the interaction between these two systems to reduce forgetting. The proposed Lifelong Self-Supervised Domain Adaptation (LLEDA) framework draws inspiration from the CLS theory and mimics the interaction between two networks: a DA network inspired by the hippocampus that quickly adjusts to changes in data distribution and an SSL network inspired by the neocortex that gradually learns domain-agnostic general representations. LLEDA’s latent replay technique facilitates communication between these two networks by reactivating and replaying the past memory latent representations to stabilize long-term generalization and retention without interfering with the previously learned information. Extensive experiments demonstrate that the proposed method outperforms several other methods resulting in a long-term adaptation while being less prone to catastrophic forgetting when transferred to new domains.

Electric vehicles (EVs) are increasingly becoming an environmental-friendly option in current transportation systems, thanks to reduced fossil fuel consumption and carbon emission. However, the more widespread adoption of EVs has been hampered by following two factors: the lack of charging infrastructure and the limited cruising range. Energy consumption estimation is crucial to address these challenges as it provides the foundations to enhance charging-station deployment, improve eco-driving behavior, and extend the EV cruising range. In this article, we propose an EV energy consumption estimation method capable of achieving accurate estimation despite insufficient EV data and ragged driving trajectories. It consists of following three distinct features: knowledge transfer from internal combustion engine/hybrid electric vehicles to EVs, segmentation-aided trajectory granularity, time-series estimation based on bidirectional recurrent neural network. Experimental evaluation shows our method outperforms other machine learning benchmark methods in estimating energy consumption on a real-world vehicle energy dataset.

Deep neural networks technique has achieved impressive performance on semantic segmentation, while its training process requires a large amount of pixel-wise labeled data. Domain adaptation, as a promising solution, can break the restriction by training the model on synthetic data, and generalizing it in real-world data. However, there is still a lack of attention paid to the imbalance problems on semantic segmentation adaptation, including the imbalance problem between 1) source and target data and 2) different classes. To solve these problems, a progressive hierarchical feature alignment method is proposed in this article. To alleviate the data imbalance problem, the network is progressively trained by the data from multisource domains, so as to obtain domain-invariant features. To address the class imbalance problem, the features are aligned hierarchically across domains. According to the experimental results, our method shows the competitive adapted segmentation performance on three benchmark datasets.

The state-of-the-art decision and planning approaches for autonomous vehicles have moved away from manually designed systems, instead focusing on the utilisation of large-scale datasets of expert demonstration via Imitation Learning (IL). In this paper, we present a comprehensive review of IL approaches, primarily for the paradigm of end-to-end based systems in autonomous vehicles. We classify the literature into three distinct categories: 1) Behavioural Cloning (BC), 2) Direct Policy Learning (DPL) and 3) Inverse Reinforcement Learning (IRL). For each of these categories, the current state-of-the-art literature is comprehensively reviewed and summarised, with future directions of research identified to facilitate the development of imitation learning based systems for end-to-end autonomous vehicles. Due to the data-intensive nature of deep learning techniques, currently available datasets and simulators for end-to-end autonomous driving are also reviewed.

Agriculture is facing severe challenges from crop stresses, threatening its sustainable development and food security. This article exploits aerial visual perception for yellow rust disease monitoring, which seamlessly integrates state-of-the-art techniques and algorithms, including unmanned aerial vehicle sensing, multispectral imaging, vegetation segmentation, and deep learning U-Net. A field experiment is designed by infecting winter wheat with yellow rust inoculum, on top of which multispectral aerial images are captured by DJI Matrice 100 equipped with RedEdge camera. After image calibration and stitching, multispectral orthomosaic is labeled for system evaluation by inspecting high-resolution RGB images taken by Parrot Anafi Drone. The merits of the developed framework drawing spectral-spatial information concurrently are demonstrated by showing improved performance over purely spectral-based classifier by the classical random forest algorithm. Moreover, various network input band combinations are tested, including three RGB bands and five selected spectral vegetation indices, by sequential forward selection strategy of wrapper algorithm.

Early driver intention prediction plays a significant role in intelligent vehicles. Drivers exhibit various driving characteristics impairing the performance of conventional algorithms using all drivers' data indiscriminatingly. This paper develops a personalized driver intention prediction system at unsignalized T intersections by seamlessly integrating clustering and classification. Polynomial regression mixture (PRM) clustering and Akaike's information criterion are applied to individual drivers trajectories for learning in-depth driving behaviors. Then, various classifiers are evaluated to link low-level vehicle states to high-level driving behaviors. CART classifier with Bayesian optimization excels others in accuracy and computation. The proposed system is validated by a real-world driving dataset. Comparative experimental results indicate that PRM clustering can discover more in-depth driving behaviors than manually defined maneuver due to its fine ability in accounting for both spatial and temporal information; the proposed framework integrating PRM clustering and CART classification provides promising intention prediction performance and is adaptive to different drivers.

Awareness of driver workload (DW) plays a paramount role in enhancing driving safety and convenience for intelligent vehicles. The DW prediction systems proposed so far learn either from individual driver's data (termed personalized system) or existing drivers' data indiscriminately (termed average system). As a result, they either do not work or lead to a limited performance for new drivers without labeled data. To this end, we develop clustering-aided approaches exploiting group characteristics of the existing drivers' data. Two clustering aided predictors are proposed. The first is clustering-aided regression (CAR) model, where the regression model for the cluster with the highest likelihood is adopted. The second is clustering-aided multiple model regression model, where the concept of multiple models is further augmented to CAR. A recent dataset from real-world driving experiments is adopted to validate the algorithms. Comparative results against the conventional average system demonstrate that by incorporating clustering information, both the proposed approaches significantly improve workload prediction performance.

Adapting in-vehicle systems (e.g., advanced driver assistance systems and in-vehicle information systems) to individual drivers' workload can enhance both safety and convenience. To make this possible, it is a prerequisite to infer driver workload so that adaptive aiding can be provided to the driver at the right time and in an appropriate manner. Rather than developing an average model for all drivers, a personalized driver workload inference (PDWI) system considering individual drivers driving characteristics is developed using machine learning techniques via easily accessed vehicle related measurements (VRMs). The proposed PDWI system comprises two stages. In offline training, individual drivers workload is first automatically splitted into different categories according to its inherent data characteristics using fuzzy C-means (FCM) clustering. Then an implicit mapping between VRMs and different levels of workload is constructed via classification algorithms. In online implementation, VRMs samples are classified into different clusters, consequently driver workload type can be successfully inferred. A recently collected dataset from real-world naturalistic driving experiments is drawn to validate the proposed PDWI system. Comparative experimental results indicate that the proposed framework integrating FCM clustering and support vector machine classifier provides a promising workload recognition performance in terms of accuracy, precision, recall, F1-score, and prediction time. The interindividual differences in term of workload are also identified and can be accommodated by the proposed framework due to its adaptiveness.