Faisal Z. Qureshi

In this work we introduce AeCSAD, a student–teacher framework designed to detect two types of anomalies in images of industrial components: structural anomalies (physical defects) and logical anomalies (incorrect relationships between components). Unlike prior methods, AeCSAD extends the component segmentation–based logical anomaly detection scheme (c. 2024) with self-attention mechanisms, enabling more effective relational modeling. We demonstrate consistent improvements on the MVTec LOCO Anomaly Detection benchmark. Specifically, AeCSAD employs a global student network with self-attention for reasoning across distant components, complemented by a local student network for fine-grained analysis. Additionally, a patch histogram module measures the frequency distribution of components, allowing the system to detect irregularities in their occurrence. During inference, anomaly scores from the histogram module and the fused local–global networks are combined to produce the final anomaly score. Experiments show that AeCSAD achieves superior average AUROC performance on both structural and logical anomaly detection tasks compared to prior approaches.

Hyperspectral images capture rich spectral information that enables per-pixel material identification; however, spectral mixing often obscures pure material signatures. To address this challenge, we propose the Latent Dirichlet Transformer Variational Autoencoder (LDVAE-T) for hyperspectral unmixing. Our model combines the global context modeling capabilities of transformer architectures with physically meaningful constraints imposed by a Dirichlet prior in the latent space. This prior naturally enforces the sum-to-one and non-negativity conditions essential for abundance estimation, thereby improving the quality of predicted mixing ratios. A key contribution of LDVAE-T is its treatment of materials as bundled endmembers, rather than relying on fixed ground truth spectra. In the proposed method our decoder predicts, for each endmember and each patch, a mean spectrum together with a structured (segmentwise) covariance that captures correlated spectral variability. Reconstructions are formed by mixing these learned bundles with Dirichlet-distributed abundances garnered from a transformer encoder, allowing the model to represent intrinsic material variability while preserving physical interpretability. We evaluate our approach on three benchmark datasets, Samson, Jasper Ridge, and HYDICE Urban and show that LDVAE-T consistently outperforms state-of-the-art models in abundance estimation and endmember extraction, as measured by root mean squared error and spectral angle distance, respectively.

Compositional Zero-Shot Learning (CZSL) is a critical task in computer vision that enables models to recognize unseen combinations of known attributes and objects during inference, addressing the combinatorial challenge of requiring training data for every possible composition. This is particularly challenging because the visual appearance of primitives is highly contextual; for example, ``small'' cats appear visually distinct from ``older'' ones, and ``wet'' cars differ significantly from ``wet'' cats. Effectively modeling this contextuality and the inherent compositionality is crucial for robust compositional zero-shot recognition. This paper presents, to our knowledge, the first comprehensive survey specifically focused on Compositional Zero-Shot Learning. We systematically review the state-of-the-art CZSL methods, introducing a taxonomy grounded in disentanglement, with four families of approaches: no explicit disentanglement, textual disentanglement, visual disentanglement, and cross-modal disentanglement. We provide a detailed comparative analysis of these methods, highlighting their core advantages and limitations in different problem settings, such as closed-world and open-world CZSL. Finally, we identify the most significant open challenges and outline promising future research directions. This survey aims to serve as a foundational resource to guide and inspire further advancements in this fascinating and important field. Papers studied in this survey with their official code are available on our github: https://github.com/ans92/Compositional-Zero-Shot-Learning

The importance of preventing failures in rotating machines has led to extensive research on diagnosis and prognosis methods based on vibration data. To achieve a high generalizability for these solutions, they need to handle high discrepancies between the data used for developing the method and the data from the target machine, a lack of historical events from the target machine and multi-component fault scenarios. Currently, state-of-the-art research has successfully found solutions targeting the first two challenges. However, these almost exclusively focus on single-component fault scenarios, meaning it is unclear how a method should be constructed considering a multi-component system. Therefore, this study constructs a framework for multi-component fault scenarios called Rotating Machinery Prognostic Framework (RoMaP) that leverages the advancements in anomaly detection, fault diagnosis and RUL prediction based on vibration data. To evaluate the potential of RoMaP, an instance based on state-of-the-art research was implemented and compared to other methods based on alternative frameworks. The results show that the instance of RoMaP had the best performance across many scenarios expected in an industrial environment, suggesting that it is a suitable approach for monitoring the health of rotating machines.

Reliable methods for bearing fault diagnosis are of great importance because they provide the possibility of preventing failures in machines. A significant challenge is developing solutions that can handle the variances in the data across different domains i.e., the operational context of the bearing and settings including noise, bearing type, rotational speed, and sampling frequency. To solve this issue, a common approach is to use transfer learning. However, most existing methods either assume that faults are available in the target domain or that only operational conditions on a single machine are changed between the source and target domain. Unfortunately, neither of these assumptions can be made in an industrial environment where there is a need to deploy the method on new machines with no historical faults. Therefore, there is a need to develop methods for bearing fault diagnosis that can provide accurate predictions on a general basis without needing access to historical faults from the target domain. To address this problem, this study develops a novel method including a knowledge-enriched standardization procedure used to lower the domain shift, feature-targeted metric learning enriched with contextual data that can be deployed to any new domain without having access to historical faults. The method is compared against state-of-the-art solutions in three cases with various tasks where it achieves the highest accuracy and displays high robustness and generalizability. The results also show a low level of false positives, a high level of fault segment recall and the possibility of giving feedback to the user about the model’s decision. The conclusion is that the method can be used on a general basis for fault diagnosis of bearings without having access to historical faults from the target domain.

Hyperspectral images record electromagnetic spectrum, and each hyperspectral pixel often stores hundreds of channels. Consequently, a hyperspectral image contains an order of magnitude more information than a similarly-sized RGB color image. Concomitant with the decreasing cost of capturing these images, there is a need to develop efficient techniques for storing, transmitting, and analyzing hyperspectral images. This paper develops a method for hyperspectral image compression using implicit neural representations where a multilayer perceptron network with sinusoidal activation functions "learns" to map pixel locations to pixel spectrum for a given hyperspectral image. This representation thus acts as a compressed encoding of this image, and the original image is reconstructed by evaluating this network at each pixel location. We introduce a sampling scheme to achieve better compression times while keeping decoding errors low. The proposed method is evaluated on four benchmarks against sixteen other schemes for hyperspectral compression, and according to the PSNR and SSIM metrics, the method developed in this paper achieves state-of-the-art compression rates at low-bit rates. Additionally, we show that the proposed sampling technique reduces encoding times.

Compositional Zero-Shot Learning (CZSL) aims to predict unknown compositions made up of attribute and object pairs. Predicting compositions unseen during training is a challenging task. We are exploring Open World Compositional Zero-Shot Learning (OW-CZSL) in this study, where our test space encompasses all potential combinations of attributes and objects. Our approach involves utilizing the self-attention mechanism between attributes and objects to achieve better generalization from seen to unseen compositions. Utilizing a self-attention mechanism facilitates the model's ability to identify relationships between attribute and objects. The similarity between the self-attended textual and visual features is subsequently calculated to generate predictions during the inference phase. The potential test space may encompass implausible object-attribute combinations arising from unrestricted attribute-object pairings. To mitigate this issue, we leverage external knowledge from ConceptNet to restrict the test space to realistic compositions. Our proposed model, Attention-based Simple Primitives (ASP), demonstrates competitive performance, achieving results comparable to the state-of-the-art.

We present a method for hyperspectral pixel unmixing. The proposed method assumes that (1) abundances can be encoded as Dirichlet distributions and (2) spectra of endmembers can be represented as multivariate Normal distributions. The method solves the problem of abundance estimation and endmember extraction within a variational autoencoder setting where a Dirichlet bottleneck layer models the abundances, and the decoder performs endmember extraction. The proposed method can also leverage transfer learning paradigm, where the model is only trained on synthetic data containing pixels that are linear combinations of one or more endmembers of interest. In this case, we retrieve endmembers (spectra) from the United States Geological Survey Spectral Library. The model thus trained can be subsequently used to perform pixel unmixing on “real data” that contains a subset of the endmembers used to generated the synthetic data. The model achieves state-of-the-art results on several benchmarks: Cuprite, Urban Hydice and Samson. We also present new synthetic dataset, OnTech-HSI-Syn-21, that can be used to study hyperspectral pixel unmixing methods. We showcase the transfer learning capabilities of the proposed model on Cuprite and OnTech-HSI-Syn-21 datasets. In summary, the proposed method can be applied for pixel unmixing a variety of domains, including agriculture, forestry, mineralogy, analysis of materials, healthcare, etc. Additionally, the proposed method eschews the need for labelled data for training by leveraging the transfer learning paradigm, where the model is trained on synthetic data generated using the endmembers present in the “real” data.