Dario ALLEGRA

Ultimo aggiornamento: 23/06/2025

I suoi interessi di ricerca riguardano l'Image Processing, Il Machine Learning e la Computer Vision applicati soprattuto all'analisi di immagini/video di alimenti e alla valorizzazione del patrimonio culturale. I suoi articoli affrontano problemi che richiedano l'utlizzo di algoritmi basati sull'imaging per poter essere risolti. Tra questi: compressione delle immagini con l'uso del re-indexing, ottimizzazione del colore, estrazione di fingerprint da documenti, riconoscimento e classificazione di immagini di cibo, stima dei valori nutrizionali da contenuti multimediali, rilevamento e rimozione di artefatti da immagini o video e altro.

Guida alle tesi di laurea

Proposte per tesi di laurea triennale o magistrale.

Sperimentali e progettuali.

Perceptive Image Compression for Interactive 3D Environments

Perceptive compression is a novel approach to image encoding where the latent representation of a picture is given by a features vector which describes the high-level components of the image. The decoding process consists of feeding such vectors to a generative model. As a consequence, the decoded image will not be an accurate reconstruction of the original one but will represent the same concepts and foreground objects. The purpose of this thesis is to explore the use of such techniques to compress textures and pictures used in virtual 3D environments, such as video game scenarios and interactive scenes. More specifically, procedural scene generation is a hopeful use case for perceptive compression as the inner variability which may be a drawback in classical image compression pipelines can be exploited to enhance the generative process.

Performance analysis of Implicit Image Compression on common user applications

Implicit Image Compression is a novel image encoding paradigm where a picture is interpreted as a function from coordinates to pixels, then overfitting a neural network to such function and compressing its parameters. The decoding is done by inferring the neural network for each pixel at the desired resolution. This approach has various advantages, such as replacing complex hand-crafted compression pipelines with generic machine learning algorithms and decoupling the image representation from its original resolution. However, codecs based on this method are still experimental and exhibit substantial encoding times, with the advantage of a straightforward and easily parallelisable decoding process. The purpose of this thesis is to implement existing decoders in everyday user applications on desktop and mobile devices and compare their computational load on common user interfaces such as web pages with traditional codecs and autoencoder-based methods.

AI-Generated media compression

AI-generated images, videos and more media have depopulated to the mass of internet users during the last years due to the release of new generative models and commercial services making use of generative AI. These data have peculiarities which are reflected in both their visual appearance and frequency spectrum. The mass spread of this kind of content makes it necessary to develop specialized compression methods that exploit their characteristics to improve compression ratios. In this thesis novel pipelines for encoding and decoding AI-generated context will be developed, analyzing the efficiency of existing codecs in such tasks and their current limitations and proposing novel techniques that better fit the specific requirements for this innovative category of data.

JPEG AI compression/recompression artifacts of Synthetic Images to counter Deepfake Detection

The proliferation of generative models (e.g., GANs, Diffusion Models) has led to the creation of highly realistic synthetic images, raising concerns about misinformation and digital forensics. To address this, researchers have developed detectors trained to distinguish between real and synthetic images, often relying on subtle artifacts left by generative processes. Neural Image Codecs (NICs), such as the recente JPEG AI standard, are increasingly used for image compression due to their superior rate-distortion performance. However, NICs introduce their own artifacts, which may overlap with those of generative models. This overlap could potentially confuse detectors, leading to false negatives or positives.

How does neural image compression affect the performance of synthetic image detectors?

Specifically, if synthetic images are compressed using NICs, do detectors still recognize them as synthetic, or do NIC artifacts obscure or mimic generative artifacts, leading to misclassification? The purpose of this thesis is to develop counter-detection methods based on recompression using LIC, and to explore possible counter-measures to distinguish NIC-compressed pristine images from synthetic ones.

Decentralized overfitted video codec

The efficiency of video-on-demand (VoD) systems is strictly linked to the compression ratio achievable, especially for most popular content. In these kind of architectures, the encoding is usually performed by the distributor, centralizing the computational efforts. However, as the resolution of videos increases and the complexity of modern codecs tends to grow with each new release, it is desirable to decentralize the computational burden instead of relying on powerful servers to encode the videos.Doing so with traditional hand-crafted encoding pipelines is difficult, as their steps are often fixed and not easy to scale. Instead, overfitted codec such as INR-based ones are inherently progressive, with each iteration bringing potential improvements to the rate-distortion ratio and being potentiaylli independent from the other ones. The purpose of this thesis is to define a decentralized and distributed video encoding system where each node is rewarded for their encoding efforts on videos, published by distribution services. This approach is similar to the one Proof-of-Work systems ran by cryptocurrencies, but in this case the work needed is not artificially generated, but productive as it contributes to the provision of more compressed videos with better quality.

Other.

Il docente è disponbile a discutere anche altre idee di tesi sperimentali/progettuali.

Compilative.

Il docente è disponibile a seguire una tesi compilativa su un qualunque argomento concordato insieme allo studente.