Vision Transformer in Industrial Visual Inspection

Abstract

Artificial intelligence as an approach to visual inspection in industrial applications has been considered for decades. Recent successes, driven by advances in deep learning, present a potential paradigm shift and have the potential to facilitate an automated visual inspection, even under complex environmental conditions. Thereby, convolutional neural networks (CNN) have been the de facto standard in deep-learning-based computer vision (CV) for the last 10 years. Recently, attention-based vision transformer architectures emerged and surpassed the performance of CNNs on benchmark datasets, regarding regular CV tasks, such as image classification, object detection, or segmentation. Nevertheless, despite their outstanding results, the application of vision transformers to real world visual inspection is sparse. We suspect that this is likely due to the assumption that they require enormous amounts of data to be effective. In this study, we evaluate this assumption. For this, we perform a systematic comparison of seven widely-used state-of-the-art CNN and transformer based architectures trained in three different use cases in the domain of visual damage assessment for railway freight car maintenance. We show that vision transformer models achieve at least equivalent performance to CNNs in industrial applications with sparse data available, and significantly surpass them in increasingly complex tasks.

1. Introduction

Industrial visual inspection (VI) in production and maintenance is under constant pressure from increasing quality requirements due to rising product demands, changing material, and cost specifications. In addition, there are changing external factors, such as new and constantly changing legal requirements, standards, and norms. Further, the risk of reputational damage from substandard products is steadily increasing due to the growing information availability and distribution through digital channels, such as social media, video sharing platforms, and review websites. Visual quality assurance is still predominantly performed or supported by human inspectors, which has several drawbacks, that have been studied by Steger et al. [1] and Sheehan et al. [2]. These include, but are not limited to, high labor cost, low efficiency, and low real-time performance on fast moving inspection objects or large surface areas. According to Swain and Guttmann [3], minimal error rates of 0.1% are reachable for very simple accept/reject tasks, which do account for negative influences of typical human inconsistency features, such as temporal fatigue. Though highly dependent on the inspection task, the findings of Drury and Fox [4] state that in more complex tasks error rates of 20% to 30% are observable. This can be seen as a rough general estimate for human error rates in visual inspection.

In order to improve efficiency and performance, as well as reduce cost, several contributions from the domain of computer vision (CV) have been proposed to automate VI. Recently, deep learning-based models, such as convolutional neural networks (CNN) [5] superseded traditional feature-based methods. In 2017, transformers, the newest type of model in deep learning, started breaking performance records in the field of natural language processing (NLP) [6,7,8,9], but also made its way into CV and claimed benchmarks there in the last two years [10,11,12]. Recently, there have already been first applications of vision transformers in VI [13,14]. While both publications demonstrate the capability of vision transformers in VI, they only cover one specific task, respectively. The work of Liu et al. [13] features a relatively simple task with small training datasets and low intra-class variation, whereas Wang et al. [14] address a more complex task, regarding number of classes, as well as detection targets per image, with a dataset of approximately twice the size. As we will discuss later, it is not possible to make a general statement about the capability of vision transformer for maintenance VI applications from their results.

To address this we selected a scenario that is a good representation of the area of maintenance VI, as it covers a wide variety of inspection use cases, to investigate the applicability of vision transformers. Our scenario comes from the context of rail freight transport, specifically, we investigate the inspection of returning freight cars after train decomposition at railroad yards to ensure operational, as well as traffic, safety. Out of the wide variety of possible use cases, we consider three learning tasks, characterized by different challenges such as high intra-class variation, large number of classes, or detection targets per image and small target objects. The first use case is damage detection in wagon sheet metal flooring with one distinct damage class that is very small in size and has similar features, as well as complexity to the one addressed by Liu et al. [13]. The second use case covers damage detection in wooden load spaces with three damage classes containing strong intra-class variation, due to different wood textures, as well as deterioration states. This one shows parallels to the work of Wang et al. [14] regarding its properties and complexity. The third use case, the localization and recognition of wagon caption characters as a foundation for matching wagons with an internal database, is characterized by a high number of classes and targets per image. This marks a complexity level beyond previous publications employing vision transformers in the context of AVI. One common challenge of all these use cases is the small amount of labeled training data for vision tasks of their complexity, typical for industrial applications. As a contribution, we evaluate how well three state-of-the-art (SOTA) vision transformer-based models, two of which, to our knowledge, have not been applied to VI before, are able to overcome the aforementioned challenges. They are compared with each other and to four established CNN models, which serve as baselines for the current SOTA in VI. We will show that performance wise the transformer architecture is on par with the CNN’s on the easier tasks, but really shows its strengths with rising task complexity.

The remainder of this paper is structured as follows: Section 2 provides an overview of related work regarding transformer models and deep-learning-based industrial VI. Section 3 introduces the learning tasks and describes the corresponding datasets. Subsequently, Section 4 describes the experiment settings and presents the obtained results. Finally, Section 5 concludes the paper with a summary of the results and discusses their implications for the direction of future research.

2. Related Work

This chapter is divided into three subsections. Section 2.1 covers the origin and evolution of transformer models in NLP. This is followed by their transfer into CV in Section 2.2 and concluded with Section 2.3 covering recent publications applying CNN, as well as vision transformers to VI use cases, on which we based our model selection.

2.1. Transformer Models in NLP

Transformers are sequence to sequence models that were proposed for the first time by Vaswani et al. in 2017 [7] in the context of natural language processing (NLP). They demonstrated exemplary performance in a broad range of language tasks, such as text classification, machine translation, and question answering [11], improving on previous records set by recurrent neural networks (RNN). This became possible through the combination of several, new or already existing, concepts. Word embeddings enable them to capture the semantic relations between words. Multi-head self-attention refines these raw embeddings in the context of specific inputs, as well as parallelizes the computation, by working on all input tokens at once, in contrast to the sequential process of RNNs. Lastly, they also utilize residual connections that increase the availability of information from shallower layers across the whole network and improve model convergence capability and speed by increasing gradient flow during backpropagation. The parallelization also makes them extremely scalable, which resulted in model size growth in popular models from 213 million parameters of the original transformer model to 340 million parameters of BERT large [8], 1.5 billion parameters of GPT-2 [15], and 175 billion parameters of GPT-3 [9]. This scaling resulted in an ever-growing performance and ability to generalize, but was not achieved without any drawbacks.

Despite their great capabilities, transformers face similar and, in some cases, amplified versions of the problems that RNNs had as well. First, like with most state-of-the-art deep learning models, large datasets are required for training. Second, they show slower convergence than, e.g., RNN or CNN models, because their structure does not inject them with prior knowledge their developers deem useful for the task at hand (inductive bias). This results in a larger solution space that is searched during training and, therefore, requires more time to converge. Third, extremely powerful hardware, regarding memory and floating-point operations per second (FLOPS), is required to train them due to the quadratic scaling of self-attention with the input length.

There have been several efforts to overcome these problems, mainly focused on reducing the complexity of the self-attention computation. Kitaev et al. [16] introduced locality sensitive hashing attention to reduce the complexity of self-attention to $O\left(nlog\right(n\left)\right)$ by approximation. Additionally, they proposed reversible layers, which enable the storage of only the final activations for backpropagation, and, therefore, reduce the memory footprint at cost of additional calculations. Wang et al. [17] improve on this with their Linformer, bringing the complexity of the approximation down to $O\left(n\right)$ in time and memory space. This is possible because the result of the softmax operation in self-attention, the context mapping matrix P, is of low rank (Equation (1)). Therefore, the self-attention result can still be well approximated with a lower dimensionality linear projection of the key and value matrix (Q, V).

$$\mathrm{Self}\text{-}\mathrm{Attention}=\underset{\mathrm{P}}{\underbrace{\mathrm{softmax}\left(\frac{Q{W}^{Q}K{W}^K}{\sqrt{d_k}}\right)}}V{W}^V$$

(1)

2.2. Transformer Models in CV

The impressive results of these models in NLP has drawn the attention of CV researchers to adapt them to their domain. To be able to feed image data into transformers it has to be transformed into sequences. Two different lines of pre-processing have been established depending on the role of the transformer model in the overall architecture: for backbone networks such as the vision transformer (ViT) [18] or the shifted windows transformer (Swin) [19] the input image is divided into patches. These are flattened and then each of them serves as one input sequence/token, like in NLP. Detection heads such as the detection transformer (DETR) [20] and the dynamic head transformer (DyHead) [21] process the flattened features extracted by an upstream CNN. In the following, we will present how DETR, deformable DETR (DDETR) [22], as well as Swin are structured and how they adapt ideas from NLP to reduce the complexity of the self-attention computation.

The transformer model in DETR is structurally almost identical to the original model introduced by Vaswani et al. The main differences are: it utilizes features extracted by a CNN backbone as input as visualized in Figure 1. These feature maps are encoded into a sequence with reduced overall dimension using a 1 × 1 convolutional layer. Static two-dimensional positional encodings are added to this sequence, as well as the input to all other encoder and decoder layers instead of only the first one. The decoder’s input is not a forward masked version of the encoder input, instead so-called object queries are used. The model features two output branches, one for class- and one for bounding box prediction. Additionally, the outputs do not require any non-maximum suppression (NMS) post-processing, because the predictions are optimized to find the best bi-partite matching with the ground truth employing the Hungarian matching algorithm [23]. Although, the performance of this model is comparable to F-RCNN’s with the same backbone, it has two problems: slow convergence speed, which was reported to be 48 days on one Nvidia V100 GPU [24], when trained from scratch; and high computational cost due to the use of full (self-)attention.

DDETR improves on DETR by proposing deformable (self-)attention. It is based on the observation that the attention matrices are very sparse usually and the results are dominated by a small number of keys for each query. Therefore, the number of keys that is considered by a query is limited to k, to reduce the number of required operations. The top center of Figure 2 shows the additional linear layer that is established to learn which k keys, around the reference point $p_q$, associated with the given query features $z_q$, to use for the attention computation. This concept, inspired by deformable convolutions [25], reduces the training time to about 14 V100 days, while also improving the performance [22]. In addition to deformable (self-)attention, Dai et al. propose three additional improvements: extension of deformable attention to multiple feature map scales, iterative bounding box refinement, where each decoder layers further refines the predictions of the previous one and a two-stage version of the model, with an additional encoder part of DDETR acting as a proposal generator for a full DDETR head.

Opposed to DETR and DDETR the Swin transformer operates on embedded, flattened patches of the input image and can function as the backbone of detection architectures like faster R-CNN (F-RCNN), Retina Net, or Yolo. To reduce the number of operations to compute self-attention, it is not evaluated over all patches, but only for non-overlapping windows containing $M\times M$ patches (Figure 3).

The window positions are not static, but shifted every layer to enable cross-window information flow, which partly preserves the global modeling power of self-attention. After a specific number of layers neighboring features are merged, by concatenating groups of $2\times 2$ patches, effectively reducing the feature resolution by half but doubling the depth. Figure 4 depicts the location of these patch merging layers found after 2, 4, 10, and 12 transformer blocks in the smallest model version. They fulfill a similar function to pooling layers and the increasing number of convolutional kernels in deeper layers in CNNs. There are four different sizes of the model: tiny, small, base, and large, reaching from 36 to 197 million parameters. All size variants of the model reduce the feature resolution to 1/16th of the input, similar to many established CNN backbones. This makes it relatively easy to exchange the backbone in many object detection or segmentation architectures with a Swin transformer.

Transformer detection heads, such as DETR, as well as backbones, such as Swin, have been shown to yield exceptional performances in object detection posting mean average precision (mAP) above 60% on the COCO benchmark dataset [26] outperforming the best CNN model YOLOR-D6 [27] by more than 2.5%.

2.3. Recent Visual Inspection Examples Utilizing Deep Learning

Since CNN’s have made their entrance to CV with AlexNets first place in the 2012 Image Net Large Scale Visual Recognition Challenge (ILSVRC) [28], they have advanced the state-of-the art in closely related areas, such as automated visual inspection, by superseding traditional feature-based methods.

Zeng et al. [29] proposed a CNN-based method for the inspection of 23 categories of train-bogey-parts, such as wheels, brakes, springs, bearings, and screws. They utilize a F-RCNN [30] architecture with ResNet-101 [31] backbone enhanced by a FPN [32] and an additional RNN submodel. This submodel is establishes a new form of gated recurrent unit (GRU) [33] to learn the strong structural correlation in technical systems. The model is evaluated on a dataset of 1130 images, containing 23 annotation classes in three sizes, where it achieves a mAP of 87.18%. In addition to this, the model is assessed on the benchmark datasets Pascal VOC 2007 and 2012 [34] and attains mAPs of 79.83% and 75.24%, respectively.

Chun et al. [35] developed a method for automatic defect detection on asphalt pavements utilizing a Yolo V3 [36] architecture. They trained their model to locate and classify four damage types: longitudinal, transverse, and alligator cracks, as well as potholes with a dataset of approximately 5000 color images and achieved an $F1\text{-}Score$ of 60%.

Very recently, one of the first VI papers applying a transformer model to crack segmentation in stone and concrete surfaces has been published by Liu et al. [13]. Inspired by SegNet, their architecture follows an encoder–decoder structure [37] with the convolutional layers in encoder and decoder replaced by self-attention blocks. They also proposed a scaling attention block to execute the processing in the feature aggregation path. They achieved new performance records on the public datasets Crack Tree 260 [38], CrackLS 315 [39], and Stone 331 [40], improving the previous results by 0.6%p to 2.1%p. Wang et al. [14] tackled the detection of major, safety critical, components of railway tracks with a transformer model. Their models employs a ResNet-50 or Darknet-53 backbone and transformer detection head, with two parallel feed forward networks, one for class and one bounding box prediction, to generate the output. They achieve a mAP of 61.9% with their best model.

Model Selection Based on Benchmarks and Recent Applications

We selected CNN and transformer models based on their benchmark results on the COCO dataset on the one hand and their occurrence in recent VI research papers on the other hand [29,35,41,42,43]. The models are DETR, DDETR and Swin, Yolo V3, F-RCNN, and Retina Net. All models except Yolo V3 and Retina Net utilized a ResNet-50 backbone. While Yolo used its original DarkNet-53 backbone, Retina Net was employed with three different ones: ResNet-50 for comparison with the other models; ResNext-101 (Retina Next) as a representative of larger more recent CNN architectures with a similar number of parameters and performance to the transformer models [44]; Swin in its “small” version to have a detection head for the only transformer backbone in this study. To complete the list we would have liked to evaluate the model by Wang et al. as well, but unfortunately it was not accessible.

3. Learning Task Description

As stated in the introduction, with this paper we want to reach a more general conclusion about the suitability and performance of vision transformers in industrial VI scenarios with small datasets. For this purpose, we have selected learning tasks from the context of rail freight transport. Specifically, we investigate the inspection of returning freight cars after train decomposition at railroad yards. The damage register contains over 1000 defect codes that cover things such as damaged load spaces, damaged or disconnected grounding cables, damage to bogie parts such as springs or wheels, but also unreadable captions or warning signs because of abrasion or graffiti contamination. Out of the wide variety of possible use cases, we consider three learning tasks, characterized by differences in intra-class variation, number of classes and detection targets per image.

The first use case, in the following referred to as sheet metal (SM) use case, covers the detection of small to medium size holes in sheet metal floorings of wagons, that cover $2\times {10}^{-2}\%$ of the inspection image on average. This is important as load, for these wagons usually bulk cargo, induced by vibrations during transport, can fall through the holes, which, on the one hand, is a commercial damage because of the lost goods and on the other hand particles or small gravel-like objects that land on the railway tracks pose a danger to traffic safety.

Figure 5 illustrates the two main challenges of this use case: On the one hand, the generally small size of the defects (left) and on the other hand the varying light conditions, as well as eventual reflections caused by wet surfaces (right).

The task is formulated as a one-class object detection problem. The dataset of this use case consists of 192 top view images, that contain a total of 394 damages, which is equal to 2.05 damages per image.

The second use case features three different damage types on wooden freight car load spaces (WLS). Its dataset is made up of 156 top view images including 255 annotations, with eight samples for damaged load support bearings or frames, 56 damaged wooden fillings of load supports, and 207 damages to floor boards (

Table 1. Class distribution of wooden load space dataset.

Damage Type	Number of Occurences
Load support bearing	8
Damaged wooden filling	56
Damaged board	206

).

The three damage types are depicted by example in Figure 6: the leftmost image shows that the connection to the bearing of a foldable load support has been detached by overstrain. In the middle one, a part of the wooden filling of a load support broke out of its frame and the one right shows a broken floor board. Broken floor boards mainly pose the threat of injury to loading or maintenance personal, when stepping on them, while damages to the load supports are influencing the proper securing of the cargo. The dataset characteristics are high color variety of the wooden components and a higher number of classes compared to sheet metal, as well as strong class imbalance.

The third use case, character recognition (CR), is the recognition of errors of captions on the side of the cars, such as loading tables, maximum load volume, length of the wagon from buffer to buffer, and other specifications. Most of these captions must be present and properly readable for the wagons to be legally in operation. Additionally, caption matching with an internal database can be the corner stone for a future system that automatically detects damage, assigns them to the database entry of the corresponding wagon and schedules its next workshop maintenance date. Figure 7 depicts some examples of captions. The base for the caption matching is the localization and classification of 43 different characters including letters, numbers, and special characters. The corresponding dataset features 200 side view greyscale images containing 21,571 bounding boxes, which is equal to 108 per image on average. Figure 8 visualizes the strongly imbalanced distribution of the 43 classes, which is one of its challenges, in addition to it featuring a lot of classes and detection targets per image.

For all three tasks, large-scale, high-definition, line-scan images with heights from 2048 to 4096 pixels and variable width are used as input data. The images were captured throughout day and night, across all seasons, at railroad yards in different locations from frames spanning over the tracks. Therefore, the light conditions vary from normal daylight to artificial lighting at night. Due to the widths of some images reaching more than 20,000 pixels, they are resized to limit their memory consumption in model training. The dimensions are adjusted so that the longer side, which is usually the width, is exactly 3072 pixels and the height is set accordingly to keep the aspect ratio constant.

Table 2. Characteristics of the datasets utilized in this study.

Task	# Images	# Annotations	# Classes	Input Size (w, h)	Step Size
Sheet metal flooring	192	394	1	(3072, *)	-
Sheet metal flooring,
windowed	219	223	1	(1024, 1024)	800, 800
Wooden load space	156	255	3	(3072, *)	-
Wood load space,
windowed	746	957	3	(512, 1024)	400, 400
Characters	200	21.571	43	(3072, *)	-
Characters,
windowed	1871	27.81	43	(1024, 1024)	800, 800

* corresponding height to keep aspect ratio.

summarizes detailed information about all datasets.

4. Experiments and Results

4.1. Experiment Settings

The experiments were executed with the complete resized images, as well as with crops generated by a sliding window approach, applied to the full resolution original images, as input. Due to the tall and narrow shapes of the damages in the wooden load space (WLS) use case, we chose a rectangular window shape of $1024\times 512$, while the sheet metal and character recognition (CR) use cases utilize a square shape of $1024\times 1024$. The crop resulting from a certain window position is only used if it contains a visible bounding box as this showed to yield the best performance. Table 2 summarizes the characteristics of the six utilized datasets. The datasets are split into training, validation, and test sets with a ratio of 80%, 10%, and 10%, respectively. We balanced the ratio of classes for WLS and CR in all sets equally so that the ratio was the same as in the original data to avoid under-representation of certain classes.

Transfer learning is utilized by initialization with weights from models pre-trained on COCO to reduce the training time and number of required samples. We selected the Adam-W optimizer [45] due to its good performance on benchmark datasets in CV [19,20,46,47,48,49]. To achieve better optimized weights due to smaller possible increments of adaptation, the initial learning rate of $1·{10}^{-4}$ is decayed by a factor of 0.1 at epoch 40 for the DETR family models and at epoch 100 for the CNN models Supplementary the gradients of DETR and DDETR are limited to $L_2\text{-}Norm\le 0.1$ during training to prevent exploding gradients. Data augmentation is used according to the original publications’ specifications [20,22,30,36,50]. This means random flipping with a probability of 50% for Retina Net and F-RCNN. DETR and DDETR utilize random flipping and random crops with resizing. For Yolo v3, a variety of photometric distortions, random crops and random flipping augment the input. Additionally, the inputs for all models are padded to a shape that makes them divisible by 32, to achieve a constant input size, which is required by all detection heads.

Our models were implemented with MMDetection, an open source object detection toolbox [51], which offers modularity for models and train/test pipelines, as well as a wide variety of pre-trained models and utility functions. We trained all models for 150 epochs on eight Nvidia V100 GPUs with 32 GB memory, with the maximum batch size the memory allows for (details in

Table 3. Model performances; bold numbers highlight the best performance for each use case.

	Model	Epoch	Batch Size	${\mathit{mAP}}_{50,\mathit{test}}$	$\mathit{F}1\text{-}{\mathit{Score}}_{\mathit{initial}}$	$\mathit{F}1\text{-}{\mathit{Score}}_{\mathit{adjusted}}$	IoU
Sheet metal	DDETR	120	1 × 8	0.699	0.640	0.704	0.25
	DETR	130	1 × 8	0.061	0.098	0.164	0.1
	F-RCNN	80	5 × 8	0.245	0.340	0.472	0.2
	Retina	100	20 × 8	0.511	0.645	0.813	0.1
	Yolo V3	120	3 × 8	0.595	0.574	0.713	0.2
	RetinaNext	150	2 × 8	0.112	0.240	0.380	0.2
	Retina Swin	130	2 × 8	0.532	0.705	0.791	0.25
Wooden load space	DDETR	110	1 × 8	0.414	0.491	0.526	0.25
	DETR	110	1 × 8	0.047	0.135	0.157	0.4
	F-RCNN	70	5 × 8	0.393	0.415	0.453	0.3
	Retina	60	7 × 8	0.389	0.475	0.644	0.2
	Yolo V3	150	2 × 8	0.522	0.531	0.571	0.3
	RetinaNext	150	2 × 8	0.056	0.195	0.244	0.4
	Retina Swin	70	2 × 8	0.335	0.407	0.556	0.1
Sheet metal window	DDETR	150	4 × 8	0.998	0.927	-	-
	DETR	130	4 × 8	0.886	0.833	-	-
	F-RCNN	120	18 × 8	0.931	0.826	0.913	0.3
	Retina	60	7 × 8	0.938	0.913	-	-
	Yolo V3	80	12 × 8	0.895	0.872	-	-
	RetinaNext	150	6 × 8	0.727	0.773	0.818	-
	Retina Swin	70	6 × 8	0.899	0.870	0.913	0.4
Wooden load space window	DDETR	150	6 × 8	0.928	0.878	0.898	0.3
	DETR	150	6 × 8	0.433	0.613	0.639	0.2
	F-RCNN	110	24 × 8	0.926	0.887	0.906	0.4
	Retina	100	28 × 8	0.942	0.913	0.922	0.4
	Yolo V3	110	14 × 8	0.862	0.779	0.802	0.4
	RetinaNext	150	6 × 8	0.498	0.759	0.802	0.1
	Retina Swin	70	6 × 8	0.949	0.907	0.927	0.3
Character recognition window	DDETR	150	6 × 8	0.676	0.779	0.912	0.2
	DETR	150	10 × 8	0.383	0.522	0.616	0.2
	F-RCNN	140	18 × 8	0.62	0.764	0.879	0.1
	Retina	70	20 × 8	0.581	0.744	0.821	0.1
	Yolo V3	120	10 × 8	0.621	0.776	0.885	0.1
	RetinaNext	150	6 × 8	0.492	0.725	0.788	0.2
	Retina Swin	90	6 × 8	0.617	0.773	0.885	0.1

). By doing so, we utilized the full capacity of the hardware, but lost some comparability between the models, because of the differing batch sizes. The loss of comparability can be justified by more stable training and better generalization as shown in the next chapter.

4.2. Results

Figure 9 shows the development of the mAP with at least 50% IoU between predicted bounding boxes and ground truth on the validation set ($mA{P}_{50,val}$) of the sheet metal use case. The strong oscillation in the performance of the transformer models is present during training of all full image use cases and makes it hard to determine if the models have converged. This instability is caused by the low batch size utilized with these models due to their high memory requirements, and further amplified by the large size of the input images.

Figure 10 visualizes the $mA{P}_{50,val}$ during training of the sliding window version of the wooden load space use case. It is clearly visible that the stability of the training process is positively influenced by the possibility to use higher batch sizes due to smaller input images. Retina Net and Retina Swin converge the fastest approximately at epoch 50, followed by DDETR, Yolo, and F-RCNN around epoch 110 and DETR at epoch 120. Retina Next seems to keep improving until epoch 150 and could possibly still benefit from continued training based on the curve’s upward gradient. This shows on the one hand how effective the improvements of DDETR on DETR are in low-data environments, and on the other hand that DDETR has a comparable convergence speed to commonly used CNN models, even with lower batch size due to high memory requirements. Retina Swin converges as fast as the fastest pure CNN in Retina Net with even better precision.

In order to determine the best weights for each model, checkpoints were saved every 10 epochs from epoch 50 onwards and evaluated on the test set. The checkpoint with best $mA{P}_{50,test}$ at the lowest epoch is selected for the following comparison. As neither recall nor precision can be neglected in industrial visual inspection, we choose their harmonic mean, also known as $F1\text{-}Score$, as the performance metric.

For the initial $F1\text{-}Score$ ($F1\text{-}Scor{e}_{initial}$) a prediction is considered to be correct, if it achieves 50% IoU and prediction confidence, which are the usual values used on benchmark datasets. Compared to benchmarks such as COCO, perfect alignment of predicted bounding boxes is not as important in real world applications as long as the targets are recognized. Therefore, the reduction in the models’ IoU threshold for the given task can improve $F1\text{-}Scor{e}_{initial}$, in many cases without drawbacks to $F1\text{-}Scor{e}_{adjusted}$. The adjusted IoU values for each model are listed in Table 3 in the last column. The prediction confidence correctness threshold is still kept at 50% to retain the same prediction quality as for $F1\text{-}Scor{e}_{initial}$. In the following, we will analyze the performance of all models summarized in Table 2, use case by use case regarding $mA{P}_{50,test}$, $F1\text{-}Scor{e}_{initial}$ and $F1\text{-}Scor{e}_{adjusted}$.

4.2.1. Full Images

In the sheet metal (SM) use case, the DDETR model achieves the highest $mA{P}_{50,test}$ of 69.9%, when looking at $F1\text{-}Scor{e}_{initial}$ it is surpassed by Retina Swin by 6.5%p. Probably, this is due to DDETRs matching loss, which optimizes the model to preferably output only one prediction per ground truth bounding box. This, in turn, leads to lower recall compared to the Retina architecture models and, therefore, to a lower F1-Score. After adjusting the IoU threshold, Retina Net takes the lead with 81.3% $F1\text{-}Scor{e}_{adjusted}$, while Retina Swin comes in second at 79.1%. The reason for this could be the fact that the weights of Retina Nets backbone and head used for transfer learning have been optimized jointly on COCO before, while the Swin weights and Retina head of Retina Swin are loaded separately before training on the SM use case. Yolo also slightly outperforms DDETR by 0.9%p. The DETR model significantly suffers from the low batch and general dataset size, as expected, resulting in the worst performance of all evaluated models in all metrics.

For the wooden load space (WLS) use case, the best performing models are Yolo V3, considering 52.2% $mA{P}_{50,test}$, respectively, 53.1% $F1\text{-}Scor{e}_{initial}$ and Retina Net with the best $F1\text{-}Scor{e}_{adjusted}$, amounting to 64.4%. This can be explained by the greater number of generated predictions before NMS of Retina Net compared to Yolo V3 (100k vs. 10k), which results in a higher recall and, consequently, a higher F1-Score. DDETR ranks third in $mA{P}_{50,test}$, second in $F1\text{-}Scor{e}_{initial}$ and third again in $F1\text{-}Scor{e}_{adjusted}$, with values of 41.4%, 49.1% and 52.6%, respectively.

Generally, the models performed worse compared to the SM use case, due to the higher difficulty caused by more damage classes, as well as higher intra-class variability. The complexity of the character recognition (CR) use case cannot be handled by any of the evaluated models, none of them were able learn to predict overlapping bounding boxes on the full images at the utilized resolution. Higher resolution training was not explored as for the DETR family models, it would have exceeded the memory capacity of our hardware.

4.2.2. Sliding Window Approach

In the sliding window version of the SM use case the DDETR outperforms the other models in all the evaluated performance metrics, followed by Retina Net in $mA{P}_{50,test}$ and Retina Net, Retina Swin, as well as F-RCNN regarding $F1\text{-}Scor{e}_{adjusted}$. The higher training batch size enables the DDETR to perform very close to 100%, while the other models do also show improvements they cannot match this. The DETR and Retina Next models also show strong improvements of their results of 30 to 40%p. There is also no performance gain in reducing the IoU threshold for DDETR, DETR, Retina Net and Yolo V3 in this use case. Compared to the SM use case with full images, the performance of all models improved significantly, because of the higher resolution input and increased pool of training samples.

Retina Swin shows the best performance considering $mA{P}_{50,test}$, $F1\text{-}Scor{e}_{adjusted}$ in the sliding window version of the WLS use case with 0.7%p and 0.5%p better results than Retina Net. This indicates that for damage characterized by a wider color variety, such as on these wooden surfaces, the global feature relations modelling capability of the Swin transformer backbone is superior to the ResNet-50. Retina Next, as well as DETR, show significant improvements over their results on full images. DETR’s convergence is still slower due to its lower inductive bias, compared to the sliding kernel of a CNN-like Retina Next, which assumes strong connections between local features.

In the sliding window version of the CR use case, DDETR bests F-RCNN considering $mA{P}_{50,test}$ by 5.5%p. Regarding $F1\text{-}Scor{e}_{initial}$, the performances are very close, differing only by 0.3%p between DETR and Yolo v3, while they are more distinct when looking at $F1\text{-}Scor{e}_{adjusted}$ with 2.7%p difference. Retina Swin’s performance is very similar to Yolo v3. The distinction between many and sometimes very similar letter shapes seems to be the strength of the DDETR detection head, possibly due to the different emphasis of the attention heads. As the inputs are greyscale images the seemingly limited feature extraction capability of the CNN backbones compared to Swin, noted in the sliding window version of the WLS use case, does not have negative impact here.

Overall, the performance of all models is significantly improved by the sliding window approach, as to be expected with higher feature resolution and larger batch sizes due to smaller inputs, at the cost of additional computation. Additionally, there is less area of the inputs that does not contain any damages, which reduces the difficulty of the task.

5. Conclusions and Outlook

In this paper, we evaluated the capability of state of the art vision transformer models for industrial VI in three representative low data use cases. These use cases are from the context of damage assessment of freight cars after train decomposition at railroad yards. The three use cases are: damage detection in sheet metal flooring, damage detection in wooden load spaces, and character localization and recognition as a foundation for matching wagons with an internal database. The last of the three marks a complexity level beyond previous publications employing vision transformers in the context of AVI.

We trained three different transformer-based architectures, DETR, DDETR, and (Retina-)Swin, of which the last two to our knowledge have not been applied to VI before. For comparison we also trained four established CNN-based architectures—F-RCNN, Retina Net, Yolo V3, and RetinaNext—that represent the current SOTA in VI as baselines. Each use case was evaluated on downscaled, full images and patches generated by a sliding window method, due to the large dimensions of the input images.

The DDETR model achieves the best results in the detection of holes in sheet metal and the recognition of characters utilizing a sliding window approach with a $F1\text{-}Scor{e}_{adjusted}$ of 92.7% and 77.9%. It seems to be very good at the differentiation between many similarly shaped objects. The Swin model with a Retina detection head delivers the best performance in the windowed version of the wooden load space use case with a $F1\text{-}Scor{e}_{adjusted}$ of 92.7%. It appears its capability to exchange information between the attention windows is especially useful when damages spread across wide areas of the images and have a high level of color variety. Considering the full image versions of the use cases the transformer models cannot surpass the CNN models, because the high memory requirements resulting from the large images limit their potential. After our studies, we can give the clear recommendation to apply vision transformer models to industrial visual inspection scenarios with typical input image sizes as they:

• Perform better than typically used CNN models;
• Show no significant difference in convergence speed compared to CNNs;
• Handle small datasets commonly utilized in industrial VI well.

Yet there is still room for improvement as the models were not modified to specifically fit the intricacies of the presented VI use cases to, e.g., better handle very large-scale input data, improve detection of small damages or become injected with prior knowledge of the typical hierarchical structure of the inspected technical systems. Based on our results it seems the combination of a Swin backbone with a DDETR detection head would achieve optimal VI results and should be evaluated in future work. Since the acquisition of labelled data is generally very expensive, other approaches are needed. Unsupervised learning offers promising concepts, such as masked patch reconstruction [52] and the soft-teacher approach [24], that need to be further explored. Solely supervised training also limits the achievable performance gains by scaling models up, which is why unsupervised pre-training played a key part in unlocking model dimensions and performances that GPT-3 exhibits in NLP, as a prime example model from the transformer family. This also enabled the concept of having one large general-purpose foundation model and use its outputs for zero- or few-shot learning of small task specific adapter models. The adaptation of this concept to VI could also prove highly valuable, as it would alleviate the labeling efforts and extremely shorten or even completely negate training time.