Hyperspectral images from newly launched (ASI-PRISMA and DLR-EnMAP) and future satellite (ESA-CHIME) are an opportunity, thanks to the high spectral resolution and full range continuity, to improve the retrieval of information about the crop parameters and status.The high dimensionality of hyperspectral data and the non-linear relationship between the crop biophysical parameters and their spectral signature make quantitative estimation of crop characteristics challenging, to address these problems we tested different configurations of neural networks (fully connected and convolutional).We tested the different architectures on two training dataset, one consists in ground data collected in three experiments, in different locations and seasons, the second one (hybrid) is composed by synthetic data generated using a radiative transfer model (PROSAIL-PRO).Preliminary results for LAI, CCC and CNC retrieval are encouraging in particular when ground data are exploited demonstrating of the potentiality of NN to fully exploit the information density of the hyperspectral data.

Leveraging state-of-the-art advancements from natural language processing (NLP), this paper pioneers the application of Transformer Neural Networks in decoding electroencephalographic (EEG) data for cognitive analysis. Harnessing the transformative power of Transformers, traditionally utilized in language understanding, this study introduces the NetTraST (Network Transformer Spatio-Temporal) model for EEG classification tasks. NetTraST integrates the paradigm-shifting capabilities of Transformer architectures with Conv1D, establishing a dual-branch architecture for extracting and processing spatial EEG features, capturing intricate neural spatial dynamics and temporal relationships within EEG sequences, illuminating nuanced temporal dynamics crucial for cognitive analysis. Transformer architecture enables the assimilation of cutting-edge techniques from the NLP domain and empowers the model to comprehend complex EEG patterns. Rigorous evaluation across diverse datasets—Thinking Out Loud, Kumar’s, and Kaneshiro’s EEG datasets—demonstrates NetTraST’s superiority over existing EEG classification methodologies not based on Transformer networks. Noteworthy advancements in accuracy substantiate the model's efficacy, surpassing previous benchmarks in multiple cognitive tasks. Furthermore, an indepth ablation study underscores the pivotal roles played by the dual branches, shedding light on their contributions to overall performance enhancement. The introduction of NetTraST marks a pivotal stride, not only harnessing the potential of Transformer architectures from NLP but also emphasizing the significance of a dual-branch approach in EEG analysis. This research's used data and implementation code is accessible at https://github.com/ignaziogallo/gallo-IJCNN2024

Recent studies have shown that Convolutional Neural Networks (CNNs) achieve impressive results in crop segmentation of Satellite Image Time-Series (SITS). However, the emergence of transformer networks in various vision tasks raises the question of whether they can outperform CNNs in crop segmentation of SITS. This paper presents a revised version of the Transformer-based Swin UNETR model adapted specifically for crop segmentation of SITS. The proposed model demonstrates significant advancements, achieving a validation accuracy of 96.14% and a test accuracy of 95.26% on the Munich dataset, surpassing the previous best results of 93.55% for validation and 92.94% for the test. Additionally, the model’s performance on the Lombardia dataset is comparable to UNet3D and superior to FPN and DeepLabV3. Experiments of this study indicate that the model will likely achieve comparable or superior accuracy to CNNs while requiring significantly less training time. These findings highlight the potential of transformer-based architectures for crop segmentation in SITS, opening new avenues for remote sensing applications.

Super-resolution is crucial in computer vision and digital image processing, aiming to enhance low-quality images’ resolution and visual quality. This paper focuses on correcting the distortion introduced by fisheye lenses and improving the resolution of images for better detail representation. Specifically, we propose an evaluation approach that benchmarks three state-of-the-art models in different categories: Real-ESRGAN (convolutions), SwinIR (transformers), and SR3 (diffusion). We evaluate their performance in super-resolution and distortion correction tasks using metrics such as PSNR and SSIM. To facilitate this evaluation, we create and release a new dataset of lunar surface images with fisheye distortion applied. Our experiments demonstrate the effectiveness of each model in handling distortion and improving image resolution. The results show that large models generally outperform medium models, and PSNR models achieve higher PSNR and SSIM scores than GAN models. Additionally, we evaluate the distortion correction by comparing the corrected images with ground truth. Our findings contribute to understanding different model categories and their performance in super-resolution and distortion correction tasks. The proposed dataset and evaluation approach can be valuable resources for future research.

Abstract A large amount of research on Convolutional Neural Networks (CNN) has focused on flat Classification in the multi-class domain. In the real world, many problems are naturally expressed as hierarchical classification problems, in which the classes to be predicted are organized in a hierarchy of classes. In this paper, we propose a new architecture for hierarchical classification, introducing a stack of deep linear layers using cross-entropy loss functions combined to a center loss function. The proposed architecture can extend any neural network model and simultaneously optimizes loss functions to discover local hierarchical class relationships and a loss function to discover global information from the whole class hierarchy while penalizing class hierarchy violations. We experimentally show that our hierarchical classifier presents advantages to the traditional classification approaches finding application in computer vision tasks. The same approach can also be applied to some CNN for text classification.

With the massive explosion of social media platforms such as Twitter and Instagram, people everyday share billions of multimedia posts, containing images and text. Typically, text in these posts is short, informal and noisy, leading to ambiguities which can be resolved using images. In this paper we will explore text-centric Named Entity Recognition task on these multimedia posts. We propose an end to end model which learns a joint representation of a text and an image. Our model extends multi-dimensional self-attention technique, where now image helps to enhance relationship between words. Experiments show that our model is capable of capturing both textual and visual contexts with greater accuracy, achieving state-of-the-art results on Twitter multimodal Named Entity Recognition dataset.

Deep metric learning plays an important role in measuring similarity through distance metrics among arbitrary group of data. MNIST dataset is typically used to measure similarity however this dataset has few seemingly similar classes, making it less effective for deep metric learning methods. In this paper, we created a new handwritten dataset named Urdu-Characters with set of classes suitable for deep metric learning. With this work, we compare the performance of two state-of-the-art deep metric learning methods i.e. Siamese and Triplet network. We show that a Triplet network is more powerful than a Siamese network. In addition, we show that the performance of a Triplet or Siamese network can be improved using most powerful underlying Convolutional Neural Network architectures.

The rise of online shopping has hurt physical retailers, which struggle to persuade customers to buy products in physical stores rather than online. Marketing flyers are a great mean to increase the visibility of physical retailers, but the unstructured offers appearing in those documents cannot be easily compared with similar online deals, making it hard for a customer to understand whether it is more convenient to order a product online or to buy it from the physical shop. In this work we tackle this problem, introducing a content extraction algorithm that automatically extracts structured data from flyers. Unlike competing approaches that mainly focus on textual content or simply analyze font type, color and text positioning, we propose a new approach that uses Convolutional Neural Networks to classify words extracted from flyers typically used in marketing materials to attract the attention of readers towards specific deals. We obtained good results and a high language and genre independence.

In this paper we introduce a novel document image classification method based on combined visual and textual information. The proposed algorithm's pipeline is inspired to the ones of other recent state-of-the-art methods which perform document image classification using Convolutional Neural Networks. The main addition of our work is the introduction of a preprocessing step embedding additional textual information into the processed document images. To do so we combine Optical Character Recognition and Natural Language Processing algorithms to extract and manipulate relevant text concepts from document images. Such textual information is then visually embedded within each document image to improve the classification results of a Convolutional Neural Network. Our experiments prove that the overall document classification accuracy of a Convolutional Neural Network trained using these text-augmented document images is considerably higher than the one achieved by a similar model trained solely on classic document images, especially when different classes of documents share similar visual characteristics.

Named Entity Recognition (NER) is an important subtask of document processing such as Information Extraction. This paper describes a NER algorithm which uses a Multi-Layer Perceptron (MLP) to find and classify entities in natural language text. In particular we use the MLP to implement a new supervised context-based NER approach called Sliding Window Neural (SWiN). The SWiN method is a good solution for domains where the documents are grammatically ill-formed and it is difficult to exploit the features derived from linguistic analysis. Experiments indicate good accuracy compared with traditional approaches and demonstrate the system's portability.

Document clustering techniques have been applied in several areas, with the Web as one of the most recent and influent. Both general-purpose and text-oriented techniques exist and can be used to cluster a collection of documents in many ways. In this work we propose an online, single-pass document clustering model that can be combined with a variety of text-oriented similarity measures. An experimental evaluation of the proposed model was conducted in the e-commerce domain. Performances were measured using a clustering-oriented metric based on F-Measure and compared with those obtained by other well-known approaches.

Many cases of remote sensing classification present complicated patterns that cannot be identified on the basis of spectral data alone, but require contextual methods that base class discrimination on the spatial relationships between the individual pixel and local and global configurations of neighboring pixels. However, the use of contextual classification is still limited by critical issues, such as complexity and problem dependency. We propose here a contextual classification strategy for object recognition in remote sensing images in an attempt to solve recognition tasks operatively. The salient characteristics of the strategy are the definition of a multiresolution feature extraction procedure exploiting human perception and the use of soft neural classification based on the multilayer perceptron model. Three experiments were conducted to evaluate the performance of the methodology, one in an easily controlled domain using synthetic images, the other two in real domains involving builtup pattern recognition in panchromatic aerial photographs and high-resolution satellite images.

The larger availability of high resolution remotely sensed data, provided by novel aircraft and space sensors, offers new perspective to image processing techniques, but it introduces also the need for operational classification tools in order to completely exploit the potentialities of these data In many applicative contexts, in particular for technological network surveillance tasks, which involve specific requirements, such as (1) high resolution and accuracy in object recognition and positioning; (2) straightforward update and change detection; (3) geographic generalisation. The application presented deals with the recognition of features of interest for the surveillance of power transmission lines using IKONOS imagery. We proposed a methodology in which multi-scale and neural techniques are synergically combined to identify features at different scales and to fuse them for class discrimination. The results obtained on a pilot area in Northern Italy proved that the combination of multi-window feature extraction and neural soft classification produced an agile and flexible model that can act as a classifier of objects that vary in shape, size and structure.