Efficient adaptive ensembling for image classification Bruno A., Moroni D., Martinelli M. In recent times, except for sporadic cases, the trend in Computer Vision is to achieve minor improvements over considerable increases in complexity.
To reverse this tendency, we propose a novel method to boost image classification performances without an increase in complexity.
To this end, we revisited ensembling, a powerful approach, not often adequately used due to its nature of increased complexity and training time, making it viable by specific design choices. First, we trained end-to-end two EfficientNet-b0 models (known to be the architecture with the best overall accuracy/complexity trade-off in image classification) on disjoint subsets of data (i.e. bagging). Then, we made an efficient adaptive ensemble by performing fine-tuning of a trainable combination layer. In this way, we were able to outperform the state-of-the-art by an average of 0.5\% on the accuracy with restrained complexity both in terms of number of parameters (by 5-60 times), and FLoating point Operations Per Second (by 10-100 times) on several major benchmark datasets, fully embracing the green AI.Source: Expert systems (Online) (2023). doi:10.1111/exsy.13424 DOI: 10.1111/exsy.13424 Metrics:
Revisiting ensembling for improving the performance of deep learning models Bruno A., Moroni D., Martinelli M. Ensembling is a very well-known strategy consisting in fusing several different models to achieve a new model for classification or regression tasks. Over the years, ensembling has been proven to provide superior performance in various contexts related to pattern recognition and artificial intelligence. Moreover, the basic ideas that are at the basis of ensembling have been a source of inspiration for the design of the most recent deep learning architectures. Indeed, a close analysis of those architectures shows that some connections among layers and groups of layers achieve effects similar to those obtainable by bagging, boosting and stacking, which are the well-known three basic approaches to ensembling. However, we argue that research has not fully leveraged the potential offered by ensembling. Indeed, this paper investigates some possible approaches to the combination of weak learners, or sub-components of weak learners, in the context of bagging. Based on previous results obtained in specific domains, we extend the approach to a reference dataset obtaining encouraging results.Source: ICPR 2022 - International Conference on Pattern Recognition, pp. 445–452, Montreal, Canada, 21-25/08/2022 DOI: 10.1007/978-3-031-37742-6_34 Metrics:
Privacy by design in systems for assisted living, personalized care and well-being: a stakeholder analysis Carboni A., Russo D., Moroni D., Barsocchi P. The concept of privacy-by-design within a system for assisted living, personalized care and well-being is crucial to protect users from misuse of the data collected about their health. Especially if the information is collected through audio-video devices, the question is even more delicate due to the nature of this data. In fact, in addition to guaranteeing a high level of privacy, it is necessary to reassure end-users about the correct use of these streams. The evolution of data analysis techniques began to take In review on an important role and increasingly defined characteristics in recent years. In this article, with reference to European projects in the AHA/AAL domain, we will see a differentiation of the concept of privacy-by-design according to different dimensions (Technical, Contextual, Business) and to the Stakeholders involved. The analysis is intended to cover technical aspects, legislative and policies-related aspects also regarding the point of view of the municipalities and aspects related to the acceptance and, therefore, to the perception of the safety of these technologies by the final end-users.Source: Frontiers in digital health (2023). doi:10.3389/fdgth.2022.934609 DOI: 10.3389/fdgth.2022.934609 Project(s): PlatformUptake.eu Metrics:
A phenotyping weeds image dataset for open scientific research Dainelli R., Martinelli M., Bruno A., Moroni D., Morelli S., Silvestri M., Ferrari E., Rocchi L., Toscano P. This in-house-built image dataset consists of 10810 weed images captured through a dedicated phenotyping activity in quasi-field conditions. The targets are seven of the most widespread and hard-to-control weeds in wheat (but also in other winter cereals) in the Mediterranean environment.
In the framework of open scientific research, our aim is to share low-cost and high-resolution images representing challenging agricultural environments where weather, lighting and other factors can change by the hour and affect the quality of images. This way the dataset could be used to train Artificial Intelligence architectures designed for weed recognition, allowing the implementation of tools directly available in the field for farmers and technicians for effective and timely weed management.
The dataset encompasses weed images ranging from the post-emergence phase (i.e. the complete cotyledons unfolding) until the pre-flowering stage. The weed selection was made by considering (i) bottom-up information and specific requests by farmers and technicians, (ii) weed susceptibility to commercial formulations for chemical control <50%, reported at least twice by field technicians, (iii) the difficulty of control considering any methods, and (iv) the type of growing season (overlapping or not with wheat). The final weeds selection encompassed both monocots (Avena sterilis and Lolium multiflorum) and dicots (Convolvulus arvensis, Fumaria officinalis, Papaver rhoeas, Veronica persica and Vicia sativa).
Image acquisition was facilitated by using a white panel as a background; this helped to (i) spread the light and thereby make the plants well-illuminated, while still avoiding strong shadows when using the flash and (ii) simplify image processing. The images were acquired with a Canon EOS 700D hand-held camera set in the macro mode with aperture, shutter speed, ISO and flash in auto mode. Photo capture timing, target distances and light conditions did not have a fixed pattern but were deliberately programmed to vary in such a way as to mimic field conditions. For image shooting at various times of the day, the only precaution was to frame the subject with homogeneous light conditions (full sunlight/full shade). The varied outdoor conditions (light, distance, timing) and camera type (RGB) with auto mode were essential features to make the images photos look similar to those that a user can take in a field, for example with a smartphone camera.
After selection and categorization, images were cropped to select the region of interest following the 1:1 ratio but maintaining a minimum size of 512 x 512 pixels.
More details on the dataset and its use for weed recognition tasks will be soon available in the proceedings of the forthcoming ECPA conference (2-6 July 2023, Bologna, Italy).
Recognition of weeds in cereals using AI architecture Dainelli R., Martinelli M., Bruno A., Moroni D., Morelli S., Silvestri M., Ferrari E., Rocchi L., Toscano P. In this study, an automatic system based on open AI architectures was developed and fed with an in-house built image dataset to recognize seven of the most widespread and hard-to-control weeds in wheat in the Mediterranean environment. A total of 10810 images were collected from the post-emergence (S1 dataset) to the pre-flowering stage (S2 dataset). A selection of pictures available from online sources (S3, 825 images) was used as a final and further independent test of the proposed recognition tool. The AI tool in the ensemble configuration achieved 100% accuracy on the validation and test set both for S1 and S2, while for S3 an accuracy of approximately 70% was achieved for weed species in the post-emergence stage.Source: ECPA 2023 - The 14th European Conference on Precision Agriculture - Unleashing the Potential of Precision Agriculture, pp. 401–407, Bologna, Italy, 2/7/2023- 6/7/2023 DOI: 10.3920/978-90-8686-947-3_49 Metrics:
Artificial intelligence for image classification of wheat's phytopathologies, pests and infestants Bruno A., Martinelli M., Moroni D., Rocchi L., Morelli S., Ferrari E., Toscano P., Dainelli R. In this work we present a solution to deal with issues related to wheat cultures: diseases, abiotic damages, weeds and pests. The core of this work is the EfficientNet architecture, in particular the b0 version which is the one with best accuracy/complexity ratio in our opinion. Results show that in this way accuracies are improved by +2% up to +10% while the complexity is unchanged. The models presented and tested during this work have been deployed and can be tested by using the mobile app Granoscan.Source: ISTI Working papers, 2023
Medical waste sorting: a computer vision approach for assisted primary sorting Bruno A., Caudai C., Leone G. R., Martinelli M., Moroni D., Crotti F. Medical waste, i.e. waste produced during medical activities in hospitals, clinics and laboratories, represents hazardous waste whose management requires special care and high costs. However, this kind of waste contains a large fraction of highly valued materials that can enter a circular economy process. To this end, in this paper, we propose a computer vision approach for assisting in the primary sorting of med- ical waste. The feasibility of our approach is demonstrated on representative datasets we collected and made available to the community.Source: IWCIM2023 - 11th International Workshop on Computational Intelligence for Multimedia Understanding, Rhodes Island, Greece, 05/06/2023 DOI: 10.1109/icasspw59220.2023.10193520 Metrics:
Efficient lung ultrasound classification Bruno A., Ignesti G., Salvetti O., Moroni D., Martinelli M. A machine learning method for classifying Lung UltraSound is here proposed to provide a point of care tool for supporting a safe, fast and accurate diagnosis, that can also be useful during a pandemic like as SARS-CoV-2. Given the advantages (e.g. safety, rapidity, portability, cost-effectiveness) provided by the ultrasound technology over other methods (e.g. X-ray, computer tomography, magnetic resonance imaging), our method was validated on the largest LUS public dataset. Focusing on both accuracy and efficiency, our solution is based on an efficient adaptive ensembling of two EfficientNet-b0 models reaching 100% of accuracy, which, to our knowledge, outperforms the previous state-of-the-art. The complexity of this solution keeps the number of parameters in the same order as an EfficientNet-b0 by adopting specific design choices that are adaptive ensembling with a combination layer, ensembling performed on the deep features, minimal ensemble only two weak models. Moreover, a visual analysis of the saliency maps on sample images of all the classes of the dataset reveals where the focus is on an inaccurate weak model versus an accurate model.Source: Bioengineering (Basel) 10 (2023). doi:10.3390/bioengineering10050555 DOI: 10.3390/bioengineering10050555 Metrics:
Efficient deep learning approach for olive disease classification Bruno A., Moroni D., Martinelli M. From ancient times olive tree cultivation has been one of the most crucial agricultural activities for Mediterranean countries. In recent years, the role of Artificial Intelligence in agriculture is increasing: its use ranges from monitoring of cultivated soil, to irrigation management, to yield prediction, to autonomous agricultural robots, to weed and pest classification and management, for example, by taking pictures using a standard smartphone or an unmanned aerial vehicle , and all this eases human work and makes it even more accessible. In this work, a method is proposed for olive disease classification, based on an adaptive ensemble of two EfficientNet-b0 models, that improves the state-of-the-art accuracy on a publicly available dataset by 1.6-2.6%. Both in terms of the number of parameters and the number of operations, our method reduces complexity roughly by 50% and 80%, respectively, that is a level not seen in at least a decade. Due to its efficiency, this method is also embeddable into a smartphone application for real-time processing.Source: ACSIS 2023 - 18th Conference on Computer Science and Intelligence Systems, pp. 889–894, Warsaw, Poland, 17-20/9/2023 DOI: 10.15439/2023f4794 Metrics:
Artificial Intelligence in TiAssisto: first results Ignesti G., Bruno A., Deri C., D'Angelo G., Salvetti O., Moroni D., Pratali L., Martinelli M. Artificial Intelligence (AI) is integrated into medical applications since its beginning. The advent of deep learning algorithms, powerful computation power and large datasets has made possible the development of numerous new medical applications. A significant part of these applications is focused on the classification or on the segmentation of medical images. In this paper we present an innovative solution for clinical images classification.Source: ISTI Working papers, 2023
Raman spectroscopy and topological machine learning for cancer grading Conti F., D'Acunto M., Caudai C., Colantonio C., Gaeta R., Moroni D., Pascali M. A. In the last decade, Raman Spectroscopy is establishing itself as a highly promising technique for the classification of tumour tissues as it allows to obtain the biochemical maps of the tissues under investigation, making it possible to observe changes among different tissues in terms of biochemical constituents (proteins, lipid structures, DNA, vitamins, and so on). In this paper, we aim to show that techniques emerging from the cross-fertilization of persistent homology and machine learning can support the classification of Raman spectra extracted from cancerous tissues for tumour grading. In more detail, topological features of Raman spectra and machine learning classifiers are trained in combination as an automatic classification pipeline in order to select the best-performing pair. The case study is the grading of chondrosarcoma in four classes: cross and leave-one-patient-out validations have been used to assess the classification accuracy of the method. The binary classification achieves a validation accuracy of 81% and a test accuracy of 90%. Moreover, the test dataset has been collected at a different time and with different equipment. Such results are achieved by a support vector classifier trained with the Betti Curve representation of the topological features extracted from the Raman spectra, and are excellent compared with the existing literature. The added value of such results is that the model for the prediction of the chondrosarcoma grading could easily be implemented in clinical practice, possibly integrated into the acquisition system.Source: Scientific reports (Nature Publishing Group) 13 (2023). doi:10.1038/s41598-023-34457-5 DOI: 10.1038/s41598-023-34457-5 Metrics:
THE D.3.2.1 - AA@THE User needs, technical requirements and specifications Pratali L., Campana M. G., Delmastro F., Di Martino F., Pescosolido L., Barsocchi P., Broccia G., Ciancia V., Gennaro C., Girolami M., Lagani G., La Rosa D., Latella D., Magrini M., Manca M., Massink M., Mattioli A., Moroni D., Palumbo F., Paradisi P., Paternò F., Santoro C., Sebastiani L., Vairo C. Deliverable D3.2.1 del progetto PNRR Ecosistemi ed innovazione - THESource: ISTI Project Report, THE, D3.2, 2023
Analysis of sea surface temperature maps via topological machine learning Conti F., Papini O., Moroni D., Pieri G., Reggiannini M., Pascali M. A. Computational methods to leverage topological features occurring in signals and images are currently one of the most innovative trends in applied mathematics. In this paper a pipeline of topological machine learning is applied to the challenging task of classifying four specific marine mesoscale patterns from remote sensing data, i.e., Sea Surface Temperature maps of the southwestern region of the Iberian Peninsula. Our preliminary study achieves an accuracy of 56% in the 4-label classification. Such results are encouraging, especially considering that the data are affected by noise and that there are low-quality/missing data. Also, the paper devises directions for future improvements.Source: ITNT 2023 - IX International Conference on Information Technology and Nanotechnology, Samara, Russia, 17-21/04/2023 DOI: 10.1109/itnt57377.2023.10139044 Project(s): NAUTILOS Metrics:
Improving plant disease classification by adaptive minimal ensembling Bruno A., Moroni D., Dainelli R., Rocchi L., Toscano P., Morelli S., Ferrari E., Martinelli M. A novel method for improving plant disease classification, a challenging and time-consuming process, is proposed. First, using as baseline EfficientNet, a recent and advanced family of architectures having an excellent accuracy/complexity trade-off, we have introduced, devised, and applied refined techniques based on transfer learning, regularization, stratification, weighted metrics, and advanced optimizers in order to achieve improved performance. Then, we go further by introducing adaptive minimal ensembling, which is a unique input to the knowledge base of the proposed solution. This represents a leap forward since it allows improving the accuracy with limited complexity using only two EfficientNet-b0 weak models, performing ensembling on feature vectors by a trainable layer instead of classic aggregation on outputs. To the best of our knowledge, such an approach to ensembling has never been used before in literature. Our method was tested on PlantVillage, a public reference dataset used for benchmarking models' performances for crop disease diagnostic, considering both its original and augmented versions. We noticeably improved the state of the art by achieving 100% accuracy in both the original and augmented datasets. Results were obtained using PyTorch to train, test, and validate the models; reproducibility is granted by providing exhaustive details, including hyperparameters used in the experimentation. A Web interface is also made publicly available to test the proposed methods.Source: Computer vision in plant phenotyping and agriculture, edited by Valerio Giuffrida, Hanno Scharr, Ian Stavness, pp. 250–263. Lausanne: Frontiers media SA, 2023
Supervised image segmentation for high dynamic range imaging Omrani A., Moroni D. Regular cameras and cell phones are able to capture limited luminosity. In terms of quality, most of the produced images by such devices are not similar to the real world. Various methods, which fall under the name of High Dynamic Range (HDR) Imaging, can be utilised to cope with this problem and produce an image with more details. However, most methods for generating an HDR image from Multi-Exposure images only focus on how to combine different exposures and do not consider the choice the best details of each image. By convers, in this research it is strived to detect the most visible areas of each image with the help of image segmentation. Two methods of producing the Ground Truth are considered, as manual and Otsu thresholding, and two similar neural networks are used to train segment these areas. Finally, it is shown that the neural network is able to segment the visible parts of pictures acceptably.Source: ICASSPW 2023 - IEEE International Conference on Acoustics, Speech, and Signal Processing, Rhodes Island, Greece, 4-10/06/2023 DOI: 10.1109/icasspw59220.2023.10193564 Metrics:
High dynamic range imaging via visual attention modules Omrani A., Moroni D. Thanks to High Dynamic Range (HDR) imaging methods, the scope of photography has seen profound changes recently. To be more specific, such methods try to reconstruct the lost luminosity of the real world caused by the limitation of regular cameras from the Low Dynamic Range (LDR) images. Additionally, although the State-Of-The-Art methods in this topic perform well, they mainly concentrate on combining different exposures and have less attention to extracting the informative parts of the images. Thus, this paper aims to introduce a new model capable of incorporating information from the most visible areas of each image extracted by a visual attention module (VAM), which is a result of a segmentation strategy. In particular, the model, based on a deep learning architecture, utilizes the extracted areas to produce the final HDR image. The results demonstrate that our method outperformed most of the State-Of-The-Art algorithms.Source: ISTI Working papers, 2023
Are we using autoencoders in a wrong way? Martino G., Moroni D., Martinelli M. Autoencoders are certainly among the most studied and used Deep Learning models: the idea behind them is to train a model in order to reconstruct the same input data. The peculiarity of these models is to compress the information through a bottleneck, creating what is called Latent Space. Autoencoders are generally used for dimensionality reduction, anomaly detection and feature extraction. These models have been extensively studied and updated, given their high simplicity and power. Examples are (i) the Denoising Autoencoder, where the model is trained to reconstruct an image from a noisy one; (ii) Sparse Autoencoder, where the bottleneck is created by a regularization term in the loss function; (iii) Variational Autoencoder, where the latent space is used to generate new consistent data. In this article, we revisited the standard training for the undercomplete Autoencoder modifying the shape of the latent space without using any explicit regularization term in the loss function. We forced the model to reconstruct not the same observation in input, but another one sampled from the same class distribution. We also explored the behaviour of the latent space in the case of reconstruction of a random sample from the whole dataset.Source: ISTI Working papers, 2023 DOI: 10.48550/arxiv.2309.01532 Metrics: