Deep convolutional neural networks (CNNs) provide an effective tool to extract complex information from images. In the area of image polarity detection, CNNs are customarily utilized in combination with transfer learning techniques to tackle a major problem: the unavailability of large sets of labeled data. Thus, polarity predictors in general exploit a pre-trained CNN as the feature extractor that in turn feeds a classification unit. While the latter unit is trained from scratch, the pre-trained CNN is subject to fine-tuning. As a result, the specific CNN architecture employed as the feature extractor strongly affects the overall performance of the model. This paper analyses state-of-the-art literature on image polarity detection and identifies the most reliable CNN architectures. Moreover, the paper provides an experimental protocol that should allow assessing the role played by the baseline architecture in the polarity detection task. Performance is evaluated in terms of both generalization abilities and computational complexity. The latter attribute becomes critical as polarity predictors, in the era of social networks, might need to be updated within hours or even minutes. In this regard, the paper gives practical hints on the advantages and disadvantages of the examined architectures both in terms of generalization and computational cost.

Compressed sensing (CS) for sensor-near vibration diagnostics represents a suitable approach for the design of network-efficient structural health monitoring systems. This article presents a solution for vibration analysis based on deep neural networks (DNNs) trained on compressed data. The envisioned maintenance system consists of a network of sensing nodes orchestrated by a very constrained centralizing unit. The latter is equipped with a microcontroller unit (MCU) that predicts the health state using the aggregated information. As a major contribution, the DNN architectures are generated automatically from the data through a procedure inspired by hardware-aware (HW) neural architecture search (NAS), called as HW-NAS-CS, which is uniquely refined with additional constraints that consider both the peculiarities of CS parameters and the limitation of embedded devices. The proposed approach has been validated using two real-world SHM datasets for vibration damage identification and eventually deployed on a low-end computing platform (the STM32L5 MCU). Results demonstrate that DNNs combined with adapted CS schemes can attain classification scores always above 90% even in case of very huge compression levels (higher than 64x): these performances significantly improve the ones attained by state-of-the-art approaches in the field, with the utmost advantage of being portable on embedded devices.

Hardware-aware neural architecture search (HW-NAS) allows the integration of convolutional neural networks (CNNs) in microcontrollers devices by automatically designing neural architectures that can fit prearranged hardware constraints. However, state-of-the-art HW-NAS target high-performance microcontrollers, whose power consumption does not meet sensing nodes requirements. This letter presents a HW-NAS generating tiny CNNs that can run on ultra-low-power microcontrollers, featuring a lightweight search procedure enabling its execution even on embedded devices. Empirical results on three well-known benchmarks for tiny computer vision proved that the proposed HW-NAS was able to generate tiny CNNs while preserving state-of-the-art classification accuracy.

Emotion recognition, among other natural language processing tasks, has greatly benefited from the use of large transformer models. Deploying these models on resource-constrained devices, however, is a major challenge due to their computational cost. In this paper, we show that the combination of large transformers, as high-quality feature extractors, and simple hardware-friendly classifiers based on linear separators can achieve competitive performance while allowing real-time inference and fast training. Various solutions including batch and Online Sequential Learning are analyzed. Additionally, our experiments show that latency and performance can be further improved via dimensionality reduction and pre-training, respectively. The resulting system is implemented on two types of edge device, namely an edge accelerator and two smartphones.

Affordance detection in computer vision allows segmenting an object into parts according to functions that those parts afford. Most solutions for affordance detection are developed in robotics using deep learning architectures that require substantial computing power. Therefore, these approaches are not convenient for application in embedded systems with limited resources. For instance, computer vision is used in smart prosthetic limbs, and in this context, affordance detection could be employed to determine the graspable segments of an object, which is a critical information for selecting a grasping strategy. This work proposes an affordance detection strategy based on hardware-aware deep learning solutions. Experimental results confirmed that the proposed solution achieves comparable accuracy with respect to the state-of-the-art approaches. In addition, the model was implemented on real-time embedded devices obtaining a high FPS rate, with limited power consumption. Finally, the experimental assessment in realistic conditions demonstrated that the developed method is robust and reliable. As a major outcome, the paper proposes and characterizes the first complete embedded solution for affordance detection in embedded devices. Such a solution could be used to substantially improve computer vision based prosthesis control but it is also highly relevant for other applications (e.g., resource-constrained robotic systems).

Thermal, electrical and mechanical stresses age the electrical insulation systems of high voltage (HV) apparatuses until the breakdown. The monitoring of the partial discharges (PDs) effectively assesses the insulation condition. PDs are both the symptoms and the causes of insulation aging and—in the long term—can lead to a breakdown, with a burdensome economic loss. This paper proposes the convolutional neural networks (CNNs) to investigate and analyze the aging process of enameled wires, thus predicting the life status of the insulation systems. The CNNs training does not require any kind of assumption of how the factors (e.g., voltage, frequency and temperature) contribute to the life model. The experiments confirm that the proposal obtains better estimations of the life status of twisted pair specimens concerning existing solutions, which are based on strong hypotheses about the life model dependency on the factors.