In this paper, we propose and evaluate a feature distillation technique for object detection under poor visibility conditions, and we analyze its impact when deployed on an FPGA platform. We demonstrate via extensive experiments how different detection architectures generalize across scenes, and we infer that a scale-permuted feature extraction is the ideal choice for detection tasks in unconstrained environments with an 11-12% gain. As verified by the experiments, image enhancement often fails to provide significant detection gains. We hence introduce a joint training in a scale-permuted student network that learns dehazed features from a dual teacher network without an explicit dehazing step. The student learns to replicate not only the teacher outputs but also the decision-making process of the teacher by using attention transfer. Although the overall goal is to produce a real-time system capable of providing driving assistance in challenging scenarios, the FPGA implementation of a scale-permuted network is the first of its kind. To achieve effective implementation of the model in FPGA technology, a high-level synthesis approach and model compression techniques are employed to obtain a deployment with a good trade-off between quality and memory footprint metrics. We develop two distilled models using the joint feature distillation technique and show that these perform better in poor visibility scenes when compared to other detectors with similar size or even bigger sizes in some cases. Our 8.5 M model shows an mAP gain of almost 1% compared to YOLOv10-M with 15 M parameters, on the Cityscapes Hazy dataset. On night images from the BDD dataset, our 8.5 M model shows an approximate mAP gain of 4% compared to YOLO26-S with 9.5 M parameters. We further perform cross-domain testing with the DriveIndia dataset to show that our models generalize well beyond the distillation distribution and can be used for generic driving scenarios.

We present a real-time, fully automatic vehicle traffic monitoring system based on phase-sensitive Optical Time Domain Reflectometry ( -OTDR) using distributed fiber optic sensors. The solution is designed for embedded on-the-edge deployment, leveraging a compact system composed of an FPGA with an integrated processor. It enables direct, in-stream processing of -OTDR signals without requiring data storage or offline analysis. The system emulates, among others facilities, the functionality of a simplified Hough Transform (HT) through a lightweight hardware architecture based on an array of binary shift registers, optimized for detecting near-vertical line segments that correspond to the signature of moving vehicles. The proposed solution demonstrates that existing optical fiber networks can be effectively repurposed for low-power, real-time traffic monitoring using dedicated embedded hardware. (Supplementary material available: https://github.com/marsistefano/ApplePie25 .)

The transformer architecture and its attention-based modules have become quite popular recently and are used for solving most computer vision tasks. However, there have been attempts to explore whether other modules can perform equally well with lower computational costs. In this paper, we introduce a nonlinear convolution structure composed of learnable polynomial and Fourier features, which allows better spectral representation with fewer parameters. The solution we propose is in principle feasible for many CNN application fields, and we present its theoretical motivation. Next, to demonstrate the performance of our architecture, and we exploit it for a paradigmatic task: image translation in driving-related scenarios such as deraining, dehazing, dark-to-bright, and night-to-day transformations. We use specific benchmark datasets for each task and standard quality parameters. The results show that our network provides acceptable or better performances when compared to transformer-based architectures, with a major reduction in the network size due to the use of such a nonlinear convolution block.

Download This Paper Open PDF in Browser Add Paper to My Library Share: Permalink Using these links will ensure access to this page indefinitely Copy URL Copy DOI

The paper focuses on the development of a vision system to automate the position control of a cultivator used for crop weeding. The vision algorithm allows monitoring of the cultivator’s misalignment with respect to crop rows, with real-time processing. The key content includes the introduction of a self-generated digital twin of the field model for numerical validation of different computer vision solutions and a comparison of three vision algorithms for measuring deviation. The objectives of the study are to improve the precision of misalignment measurements and ensure safe and accurate movement of the cultivator. The rationale behind the study is to address constraints such as camera installation and crop color, and to emphasize the importance of a confidence estimation feature for accurate measurement. The paper also provides an overview of related works in the literature, highlighting the two phases of plant identification and deviation measurement. Tests carried out on soybean and maize crops demonstrate the improvements allowed by the proposed algorithm in terms of higher measurement precision, even in the presence of high weed infestation or a significant number of missing plants. Additionally, the paper suggests analysis simplifications to enhance the algorithm’s speed while maintaining satisfactory measurement accuracy.

In precision agriculture, effective pest control helps to reduce yield loss and pesticide application. In this research, the pest to be detected and controlled is the moth lobesia botrana, which mainly attacks the vineyard. We present an automatic pest classifier based on machine learning, considering resource-constrained devices in IoT systems. Transfer learning and an ensemble of compression techniques are used to reduce the size of the classifier with a good trade-off between efficiency, effectiveness, and resource utilization. This procedure allows the achievement of a fully on-chip deployment in two technologies: esp32 and SoC-based FPGA Xilinx PYNQ-Z1 and KRIA.

In this paper, a method for estimating the blur regions of an image is first proposed, resorting to a mixture of linear and nonlinear convolutional kernels. The blur map obtained is then utilized to enhance images such that the enhancement strength is an inverse function of the amount of measured blur. The blur map can also be used for tasks such as attention-based object classification, low light image enhancement, and more. A CNN architecture is trained with nonlinear upsampling layers using a standard blur detection benchmark dataset, with the help of blur target maps. Further, it is proposed to use the same architecture to build maps of areas affected by the typical JPEG artifacts, ringing and blockiness. The blur map and the artifact map pair permit to build an activation map for the enhancement of a (possibly JPEG compressed) image. Extensive experiments on standard test images verify the quality of the maps obtained using the algorithm and their effectiveness in locally controlling the enhancement, for superior perceptual quality. Last but not least, the computation time for generating these maps is much lower than the one of other comparable algorithms.

This paper introduces NLCMap, a framework for the mapping space exploration targeting Non-Linear Convolutional Networks (NLCNs). NLCNs [1] are a novel neural network model that improves performances in certain computer vision applications by introducing a non-linearity in the weights computation. NLCNs are more challenging to efficiently map onto hardware accelerators if compared to traditional Convolutional Neural Networks (CNNs), due to data dependencies and additional computations. To this aim, we propose NLCMap, a framework that, given an NLC layer and a generic hardware accelerator with a certain on-chip memory budget, finds the optimal mapping that minimizes the accesses to the off-chip memory, which are often the critical aspect in CNNs acceleration.

Phytosanitary treatment is one of the most critical operations in vineyard management. Ideally, the spraying system should treat only the canopy, avoiding drift, leakage and wasting of product where leaves are not present: variable rate distribution can be a successful approach, allowing the minimization of losses and improving economic as well as environmental performances. The target of this paper is to realize a smart control system to spray phytosanitary treatment just on the leaves, optimizing the overall costs/benefits ratio. Four different optical-based systems for leaf recognition are analyzed, and their performances are compared using a synthetic vineyard model. In the paper, we consider the usage of three well-established methods (infrared barriers, LIDAR 2-D and stereoscopic cameras), and we compare them with an innovative low-cost real-time solution based on a suitable computer vision algorithm that uses a simple monocular camera as input. The proposed algorithm, analyzing the sequence of input frames and exploiting the parallax property, estimates the depth map and eventually reconstructs the profile of the vineyard’s row to be treated. Finally, the performances obtained by the new method are evaluated and compared with those of the other methods on a well-controlled artificial environment resembling an actual vineyard setup while traveling at standard tractor forward speed.

This paper proposes a new neural network structure for image processing whose convolutional layers, instead of using kernels with fixed coefficients, use space-variant coefficients. The adoption of this strategy allows the system to adapt its behavior according to the spatial characteristics of the input data. This type of layers performs, as we demonstrate, a non-linear transfer function. The features generated by these layers, compared to the ones generated by canonical CNN layers, are more complex and more suitable to fit to the local characteristics of the images. Networks composed by these non-linear layers offer performance comparable with or superior to the ones which use canonical Convolutional Networks, using fewer layers and a significantly lower number of features. Several applications of these newly conceived networks to classical image-processing problems are analyzed. In particular, we consider: Single-Image Super-Resolution (SISR), Edge-Preserving Smoothing (EPS), Noise Removal (NR), and JPEG artifacts removal (JAR).

This work focuses on determining the velocity profile of a granular flow at the outlet of a silo, using artificial vision techniques. The developed algorithm performs a frame enhancement through neural networks and the particle image velocimetry detects seed motion in the hopper. We process 50, 100, 150 and 200 frames of a video discharge for three different grains using: CPU and PYNQ-Z1 implementations with a simple image processing at pre-processing level, and CPU implementation using neural network. Execution times are measured and the differences between the involved technologies are discussed.

Marsi, Stefano Carrato, Sergio De Bortoli, Luca Gallina, Paolo Guzzi, Francesco Ramponi, GiovanniThis paper proposes a method for depth estimation in video sequences acquired by a monocular camera mounted on a mobile platform. The proposed algorithm is able to estimate in real time the relative distances of the objects in the field of view exploiting the parallax effect, provided the platform movement complies with a few constraints. The developed system is designed to operate at the input pixel cadence and is thus applicable to any video resolution. The final architecture, using operators no more complex than an adder and a memory that is just a fraction of a frame memory, can be realized in a low-cost FPGA.

Nowadays, the trend of the latest research in face recognition model shows that “the complex—the better” paradigm can be directly applied to these systems, whose accuracy effectively depends on both a large number of well-trained parameters and a complex functional structure. If this approach is sustainable for an offline processing on a consumer PC, it is far less appealing in the mobile environment, where processing power, as well as a high amount of onboard RAM could not be available. The distillation technique, applied on the cumbersome dlib-resnet-v1 face recognition model results in a lighter version that, while maintaining a comparable accuracy, can achieve a faster processing rate (>10×) and a lower memory occupation (1/6). The final model has been implemented on a single board PC, also using a neural hardware accelerator.

Face recognition functions are today exploited through biometric sensors in many applications, from extended security systems to inclusion devices; deep neural network methods are reaching in this field stunning performances. The main limitation of the deep learning approach is an inconvenient relation between the accuracy of the results and the needed computing power. When a personal device is employed, in particular, many algorithms require a cloud computing approach to achieve the expected performances; other algorithms adopt models that are simple by design. A third viable option consists of model (oracle) distillation. This is the most intriguing among the compression techniques since it permits to devise of the minimal structure that will enforce the same I/O relation as the original model. In this paper, a distillation technique is applied to a complex model, enabling the introduction of fast state-of-the-art recognition capabilities on a low-end hardware face recognition sensor module. Two distilled models are presented in this contribution: the former can be directly used in place of the original oracle, while the latter incarnates better the end-to-end approach, removing the need for a separate alignment procedure. The presented biometric systems are examined on the two problems of face verification and face recognition in an open set by using well-agreed training/testing methodologies and datasets.

In face recognition systems, the use of convolutional neural networks (CNNs) permits to achieve good accuracy performances, which derive largely from a huge number of well-trained parameters. While using online services any mobile device can suffice for an accurate identification, in the offline scenario, implemented on a wearable mobile hardware, it is difficult to achieve both real-time responsiveness and high accuracy. In this paper we present a solution to replace a large open source face recognizer network (provided as part of the dlib libraries), distilling its learned knowledge into a less demanding CNN. The former is used as an expert oracle that provides the targets, while the latter is trained on the same input image, following a regression approach. In addition to lightness, our CNN is trained to use smaller input images, naturally allowing the recognition of identities in a wider distance range and with a reduced amount of computation. This eventually permits the porting of the network into a dedicated mobile accelerating hardware. The hypothesis we want to demonstrate is that since the feature space topology has been deeply explored during the training of the expert network, and due to the fact that no information is created during the up sampling of a tiny face to the input size of the expert oracle, the smaller network can provide the same accuracy at a reduced computational cost.

Face recognition systems areMarsi, Stefano of great interest in manyDe Bortoli, Luca applications. We present someGuzzi, Francesco results from a comparison on different classification methods using an open source tool that works withBhattacharya, Jhilik Convolutional Neural NetworksCicala, Francesco to extract facial features. This work focuses on theCarrato, Sergio performance obtainable from a multi-class classifier, trainedCanziani, Alfredo with a reduced number images, to identify a person between aRamponi, Giovanni group of known and unknown subjects . The overall system has been implemented in an Odroid XU-4 Platform.

We present some results from an ongoing project about face detection and recognition in an apparatus wearable by a visually impaired person. Specifically, we explore the usable equipment and we experiment on the realization of three prototypes that give the opportunity of dealing with different topics, ranging from the architecture of the network to database creation, from the reliability of the identification results to real-time operation issues.

We propose a strategy to realize an innovative FPGA-based architecture able to speed up the Lucy-Richardson algorithm (LRA) for space-variant image deconvolution. The architecture exploits the possibility to distribute data into different memory blocks in the FPGA. In such a way, the algorithm execution is split into several channels operating in parallel. Since the LRA is implemented via an iterative and space-variant convolution, the approach adopted in this paper can be exploited in other similar image processing algorithms.

In this paper, we analyze the accomplishment of a study for a low cost system, named PicoAgri, to monitor the status of agricultural fields. We here deal with the small detection system we are developing representing the first element to build the final PicoAgri concept: a system consisting of two elements, an array of small satellites coupled to drones. The array of CubeSat small satellites will provide data for an initial multispectral analysis with a resolution at ground of 30 m; when required a well localized area will be explored employing suitable drones capable of analyzing the territory with a resolution at least two orders of magnitude higher.