add abstract, finish the alternatvie methods and fix some todos and improve sources

experimentalresults.typ

@@ -64,7 +64,7 @@ Which is an result that is unexpected (since one can think more samples perform
 Clearly all four graphs show that the performance decreases with an increasing number of good samples.
 So the conclusion is that the Few-Shot learner should always be trained with as balanced classes as possible.
 
-== How does the 3 (ResNet, CAML, pmf) methods perform in only detecting the anomaly class?
+== How does the 3 (ResNet, CAML, P>M>F) methods perform in only detecting the anomaly class?
 _How much does the performance improve if only detecting an anomaly or not?
 How does it compare to PatchCore and EfficientAD#todo[Maybe remove comparion?]?_
 

implementation.typ

@@ -7,15 +7,19 @@
 The three methods described (ResNet50, CAML, P>M>F) were implemented in a Jupyter notebook and compared to each other.
 
 == Experiments <experiments>
-For all of the three methods we test the following use-cases:#todo[maybe write more to each test]
+For all of the three methods we test the following use-cases:
 - Detection of anomaly class (1,3,5 shots)
+  - Every faulty class and the good class is detected.
 - 2 Way classification (1,3,5 shots)
+  - Only faulty or not faulty is detected. All the samples of the faulty classes are treated as a single class.
 - Detect only anomaly classes (1,3,5 shots)
+  - Similar to the first test but without the good class. Only faulty classes are detected.
 - Inbalanced 2 Way classification (5,10,15,30 good shots, 5 bad shots)
-- Inbalanced target class prediction (5,10,15,30 good shots, 5 bad shots)
+  - Similar to the 2 way classification but with an inbalanced number of good shots.
-
+- Inbalanced target class prediction (5,10,15,30 good shots, 5 bad shots)#todo[Avoid bullet points and write flow text?]
-Those experiments were conducted on the MVTEC AD dataset on the bottle and cable classes.
+  - Detect only the faulty classes without the good classed with an inbalanced number of shots.
 
+All those experiments were conducted on the MVTEC AD dataset on the bottle and cable classes.
 
 == Experiment Setup
 All the experiments were done on the bottle and cable classes of the MVTEC AD dataset.
@@ -23,20 +27,21 @@ The correspoinding number of shots were randomly selected from the dataset.
 The rest of the images was used to test the model and measure the accuracy.
 #todo[Maybe add real number of samples per classes]
 
-== ResNet50
+== ResNet50 <resnet50impl>
 === Approach
 The simplest approach is to use a pre-trained ResNet50 model as a feature extractor.
 From both the support and query set the features are extracted to get a downprojected representation of the images.
 The support set embeddings are compared to the query set embeddings.
 To predict the class of a query the class with the smallest distance to the support embedding is chosen.
 If there are more than one support embedding within the same class the mean of those embeddings is used (class center).
-This approach is similar to a prototypical network @snell2017prototypicalnetworksfewshotlearning.
+This approach is similar to a prototypical network @snell2017prototypicalnetworksfewshotlearning and the work of _Just Use a Library of Pre-trained Feature
+Extractors and a Simple Classifier_ @chowdhury2021fewshotimageclassificationjust but just with a simple distance metric instead of a neural net.
 
 In this bachelor thesis a pre-trained ResNet50 (IMAGENET1K_V2) pytorch model was used.
 It is pretrained on the imagenet dataset and has 50 residual layers.
 
 To get the embeddings the last layer of the model was removed and the output of the second last layer was used as embedding output.
-In the following diagram the ResNet50 architecture is visualized and the cut-point is marked.
+In the following diagram the ResNet50 architecture is visualized and the cut-point is marked.~@chowdhury2021fewshotimageclassificationjust
 
 #diagram(
   spacing: (5mm, 5mm),

introduction.typ

@@ -5,12 +5,13 @@
 Anomaly detection has especially in the industrial and automotive field essential importance.
 Lots of assembly lines need visual inspection to find errors often with the help of camera systems.
 Machine learning helped the field to advance a lot in the past.
-Most of the time the error rate is sub $.1%$ and therefore plenty of good data is available and the data is heavily unbalaned.
+Most of the time the error rate is sub $.1%$ and therefore plenty of good data and almost no faulty data is available.
+So the train data is heavily unbalaned.#cite(<parnami2022learningexamplessummaryapproaches>)
 
 PatchCore and EfficientAD are state of the art algorithms trained only on good data and then detect anomalies within unseen (but similar) data.
 One of their problems is the need of lots of training data and time to train.
 Moreover a slight change of the camera position or the lighting conditions can lead to a complete retraining of the model.
-Few-Shot learning might be a suitable alternative with hugely lowered train times and fast adaption to new conditions.
+Few-Shot learning might be a suitable alternative with hugely lowered train times and fast adaption to new conditions.~#cite(<efficientADpaper>)#cite(<patchcorepaper>)#cite(<parnami2022learningexamplessummaryapproaches>)
 
 In this thesis the performance of 3 Few-Shot learning algorithms will be compared in the field of anomaly detection.
 Moreover, few-shot learning might be able not only to detect anomalies but also to detect the anomaly class.

main.typ

@@ -52,7 +52,18 @@
   place-of-submission: "Linz",
   title: "Few shot learning for anomaly detection",
   abstract-en: [//max. 250 words
-#lorem(200)  ],
+    This thesis explores the application of Few-Shot Learning (FSL) in anomaly detection, a critical area in industrial and automotive domains requiring robust and efficient algorithms for identifying defects.
+    Traditional methods, such as PatchCore and EfficientAD, achieve high accuracy but often demand extensive training data and are sensitive to environmental changes, necessitating frequent retraining.
+    FSL offers a promising alternative by enabling models to generalize effectively from minimal samples, thus reducing training time and adaptation overhead.
+
+    The study evaluates three FSL methods—ResNet50, P>M>F, and CAML—using the MVTec AD dataset.
+    Experiments focus on tasks such as anomaly detection, class imbalance handling, and comparison of distance metrics.
+    Results indicate that while FSL methods trail behind state-of-the-art algorithms in detecting anomalies, they excel in classifying anomaly types, showcasing potential in scenarios requiring detailed defect identification.
+    Among the tested approaches, P>M>F emerged as the most robust, demonstrating superior accuracy across various settings.
+
+    This research underscores the limitations and niche applicability of FSL in anomaly detection, advocating its integration with established algorithms for enhanced performance.
+    Future work should address the scalability and domain-specific adaptability of FSL techniques to broaden their utility in industrial applications.
+  ],
   abstract-de: none,// or specify the abbstract_de in a container []
   acknowledgements: none,//acknowledgements: none // if you are self-made
   show-title-in-header: false,

materialandmethods.typ

@@ -73,24 +73,23 @@ So many more defect classes are already an indication that a classification task
 === Few-Shot Learning
 Few-Shot learning is a subfield of machine-learning which aims to train a classification-model with just a few or no samples at all.
 In contrast to traditional supervised learning where a huge amount of labeled data is required is to generalize well to unseen data.
-So the model is prone to overfitting to the few training samples.
+So the model is prone to overfitting to the few training samples.#cite(<parnami2022learningexamplessummaryapproaches>)
 
 Typically a few-shot leaning task consists of a support and query set.
 Where the support-set contains the training data and the query set the evaluation data for real world evaluation.
 A common way to format a few-shot leaning problem is using n-way k-shot notation.
-For Example 3 target classeas and 5 samples per class for training might be a 3-way 5-shot few-shot classification problem.
+For Example 3 target classeas and 5 samples per class for training might be a 3-way 5-shot few-shot classification problem.#cite(<snell2017prototypicalnetworksfewshotlearning>)#cite(<patchcorepaper>)
 
 A classical example of how such a model might work is a prototypical network.
-These models learn a representation of each class and classify new examples based on proximity to these representations in an embedding space.
+These models learn a representation of each class and classify new examples based on proximity to these representations in an embedding space.#cite(<snell2017prototypicalnetworksfewshotlearning>)
 
 #figure(
   image("rsc/prototype_fewshot_v3.png", width: 60%),
   caption: [Prototypical network for few-shots. #cite(<snell2017prototypicalnetworksfewshotlearning>)],
 ) <prototypefewshot>
 
-The first and easiest method of this bachelor thesis uses a simple ResNet to calucalte those embeddings and is basically a simple prototypical netowrk.
+The first and easiest method of this bachelor thesis uses a simple ResNet to calucalte those embeddings and is basically a simple prototypical network.
-See #todo[link to this section]
+See @resnet50impl.~#cite(<chowdhury2021fewshotimageclassificationjust>)
-#todo[proper source]
 
 === Generalisation from few samples
 
@@ -373,7 +372,7 @@ Its use of frozen pre-trained feature extractors is key to avoiding overfitting
 
 == Alternative Methods
 
-There are several alternative methods to few-shot learning which are not used in this bachelor thesis.
+There are several alternative methods to few-shot learning as well as to anomaly detection which are not used in this bachelor thesis.
 Either they performed worse on benchmarks compared to the used methods or they were released after my initial literature research.
 
 === SgVA-CLIP (Semantic-guided Visual Adapting CLIP)
@@ -393,18 +392,58 @@ If the pre-trained model lacks relevant information for the task, SgVA-CLIP migh
 This might be a no-go for anomaly detection tasks because the images in such tasks are often very task-specific and not covered by general pre-trained models.
 Also, fine-tuning the model can require considerable computational resources, which might be a limitation in some cases.~#cite(<peng2023sgvaclipsemanticguidedvisualadapting>)
 
-=== TRIDENT
+=== TRIDENT (Transductive Decoupled Variational Inference for Few-Shot Classification)
 // https://arxiv.org/pdf/2208.10559v1
 // https://arxiv.org/abs/2208.10559v1
 
-== SOT
+TRIDENT, a variational infernce network, is a few-shot learning approach which decouples image representation into semantic and label-specific latent variables.
+Semantic attributes contain context or stylistic information, while label-specific attributes focus on the characteristics crucial for classification.
+By decoupling these parts TRIDENT enhances the networks ability to generalize effectively from unseen data.~#cite(<singh2022transductivedecoupledvariationalinference>)
+
+To further improve the discriminative performance of the model, it incorporates a transductive feature extraction module named AttFEX (Attention-based Feature Extraction).
+This feature extractor dynamically aligns features from both the support and the query set, promoting task-specific embeddings.~#cite(<singh2022transductivedecoupledvariationalinference>)
+
+This model is specifically designed for few-shot classification tasks but might also work well for anomaly detection.
+Its ability to isolate critical features while droping irellevant context aligns with requirements needed for anomaly detection.
+
+=== SOT (Self-Optimal-Transport Feature Transform)
 // https://arxiv.org/pdf/2204.03065v1
 // https://arxiv.org/abs/2204.03065v1
 
+The Self-Optimal-Transport (SOT) Feature Transform is designed to enhance feature sets for tsks like matching, grouping or classification by re-embedding feature representations.
+This transform processes features as a set instead of using them individually.
+This creates context-aware representations.
+SOT can catch direct as well as indirect similarities between features which makes it suitable for tasks like few-shot learning or clustering.~#cite(<shalam2022selfoptimaltransportfeaturetransform>)
+
+SOT uses a transport plan matrix derived from optimal transport theory to redefine feature relations.
+This includes calculating pairwaise similarities (e.g. cosine similarities) between features and solving a min-cost max-flow problem to find an optimal match between features.
+This results in an doubly stochastic matrix where each row represents the re-embedding of the corresponding feature in context with others.~#cite(<shalam2022selfoptimaltransportfeaturetransform>)
+
+The transform features parameterless-ness, which makes it easy to integrate into existing machine-learning pipelines.
+It is differentiable which allows for end-to-end training. For example (re-)train the hosting network to adopt to SOT.
+SOT is equivariant, which means that the transform is invariant to the order of the input features.~#cite(<shalam2022selfoptimaltransportfeaturetransform>)
+
+The improvements of SOT over traditional feature transforms dpeend on the used backbone network and the task.
+But in most cases it outperforms state-of-the-art methods and could be used as a drop-in replacement for existing feature transforms.~#cite(<shalam2022selfoptimaltransportfeaturetransform>)
+
 // anomaly detect
-== GLASS
+=== GLASS (Global and Local Anomaly co-Synthesis Strategy)
 // https://arxiv.org/pdf/2407.09359v1
 // https://arxiv.org/abs/2407.09359v1
 
+GLASS (Global and Local Anomaly co-Synthesis Strategy) is a anomaly detection method for industrial applications.
+It is a unified network which uses two different strategies to detect anomalies which are then combined.
+The first one is Global Anomaly Synthesis (GAS), it operates on the feature level.
+It uses a gaussian noise, guided by gradient ascent and constrained by truncated projection to generate anomalies close to the distribution for the normal features.
+This helps the detection of weak defects.
+The second strategy is Local Anomaly Synthesis (LAS), it operates on the image level.
+This strategy overlays textures onto normal images using masks derived from noise patterns.
+LAS creates strong anomalies which are further away from the normal sample distribution.
+This adds diversity to the synthesized anomalies.~#cite(<chen2024unifiedanomalysynthesisstrategy>)
+
+GLASS combines GAS and LAS to improve anomaly detection and localization by synthesizing anomalies near and far from the normal distribution.
+Experiments show that GLASS is very effective and outperforms some state-of-the-art methods on the MVTec AD dataset such as PatchCore in some cases.~#cite(<chen2024unifiedanomalysynthesisstrategy>)
+
+//=== HETMM (Hard-normal Example-aware Template Mutual Matching)
 // https://arxiv.org/pdf/2303.16191v5
 // https://arxiv.org/abs/2303.16191v5

sources.bib

@@ -147,3 +147,53 @@
       primaryClass={cs.CV},
       url={https://arxiv.org/abs/2211.16191},
 }
+
+@misc{singh2022transductivedecoupledvariationalinference,
+      title={Transductive Decoupled Variational Inference for Few-Shot Classification},
+      author={Anuj Singh and Hadi Jamali-Rad},
+      year={2022},
+      eprint={2208.10559},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2208.10559},
+}
+
+@misc{chen2024unifiedanomalysynthesisstrategy,
+      title={A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization},
+      author={Qiyu Chen and Huiyuan Luo and Chengkan Lv and Zhengtao Zhang},
+      year={2024},
+      eprint={2407.09359},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2407.09359},
+}
+
+@misc{shalam2022selfoptimaltransportfeaturetransform,
+      title={The Self-Optimal-Transport Feature Transform},
+      author={Daniel Shalam and Simon Korman},
+      year={2022},
+      eprint={2204.03065},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2204.03065},
+}
+
+@misc{parnami2022learningexamplessummaryapproaches,
+      title={Learning from Few Examples: A Summary of Approaches to Few-Shot Learning},
+      author={Archit Parnami and Minwoo Lee},
+      year={2022},
+      eprint={2203.04291},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2203.04291},
+}
+
+@misc{chowdhury2021fewshotimageclassificationjust,
+      title={Few-shot Image Classification: Just Use a Library of Pre-trained Feature Extractors and a Simple Classifier},
+      author={Arkabandhu Chowdhury and Mingchao Jiang and Swarat Chaudhuri and Chris Jermaine},
+      year={2021},
+      eprint={2101.00562},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2101.00562},
+}