Current students


STUCCHI DIEGOCycle: XXXV

Section: Computer Science and Engineering
Advisor: BORACCHI GIACOMO
Tutor: GATTI NICOLA

Major Research topic:
MULTIMODAL AND REFERENCE-BASED ANOMALY DETECTION ALGORITHMS

Abstract:
Anomaly Detection (AD) is a fundamental problem both in the industry (e.g. high-end manufacturing) and in many Computer Vision applications. Indeed, visual inspection from an expert is time consuming, expensive or not repeatable, hence not always a viable option. Moreover, algorithms for automated AD are essential to guarantee a full control in high throughput applications. Often, the AD context includes also reference images representing the anomaly-free subject to analyze. My research focuses on the design of reference-based AD algorithms in scenarios where the sensed data and the reference cannot be directly compared due to noise, misregistration and -- most remarkably -- inconsistency of their modalities. There are two main challenges: solving the geometric-photometric misalignment and designing a statistical decision rule (fault / no fault, change / no change, anomalous / normal).
To tackle the former, I worked in a collaboration with Tampere University on an algorithm recovering a multimodal image from noisy linear measurements of distinct misaligned modalities. The designed solution is a multi-scale iterative algorithm consisting in an affine transformation fitting and a polynomial intensity transformation estimation. Although we implemented a first working solution of this algorithm, the collaboration is still ongoing to produce an improved final version of the implementation.
On the other hand, I worked on a statistical tool for batch-wise Change Detection, called Multimodal QuantTree. MMQT is a density-based method for batch-wise CD that learns a model from a training set of stationary batches of data, where each batch might be drawn from a different distribution or modality. The algorithm uses a single histogram and a clustering step to estimate the number of different modalities characterizing the data and detects as non-stationary batches whose statistic exceeds a set threshold. We submitted a paper to ICPR (the decision is due on September 30th).
Furthermore, I worked on the design of a light-weight Fault Impact Estimation algorithm based on a feed-forward fully-connected neural network that estimates the impact of a fault affecting the output of an image processing application. In this scenario, the multimodality consists in the different spatial resolution of the considered satellite images and in their different nature (artificial vs. man-made scenarios). A remarkable property of the method is that its computational cost is significantly lower than the cost of the replication-based methods considered as a baseline. In fact, FIE raises an alarm and triggers the recomputation only when the detected fault heavily affects the output image. Also, we developed a transfer learning method to address images coming from different modalities (i.e. spatial resolutions) without requiring a new training process. We submitted a paper to IEEE Transaction on Computers and we are currently writing a revised version.
Currently, I am collaborating with EPTA, the industrial sponsor of my PhD scholarship. In this industrial scenario, the objective is to design a reference-based Anomaly Detection algorithm able to detect malfunctions in the refrigerated cabinets. This scenario falls in the AD framework of interest, since EPTA provided both CFD thermo-dynamic simulations describing the temperatures of an operating cabinet (reference images) and IR thermal low-resolution misaligned acquisitions (sensed images) against which to compare the reference.