This thesis investigates deep perceptual loss and (deep perceptual) similarity; methods for computing loss and similarity for images as the distance between the deep features extracted from neural networks. The primary contributions of the thesis consist of (i) aggregating much of the existing research on deep perceptual loss and similarity, and (ii) presenting novel research into understanding and improving the methods. This novel research provides insight into how to implement the methods for a given task, their strengths and weaknesses, how to mitigate those weaknesses, and if these methods can handle the inherent ambiguity of similarity.

Society increasingly relies on computer vision technology, from everyday smartphone applications to legacy industries like agriculture and mining. Much of that groundbreaking computer vision technology relies on machine learning methods for their success. In turn, the most successful machine learning methods rely on the ability to compute the similarity of instances.

In computer vision, computation of image similarity often strives to mimic human perception, called perceptual similarity. Deep perceptual similarity has proven effective for this purpose and achieves state-of-the-art performance. Furthermore, this method has been used for loss calculation when training machine learning models with impressive results in various computer vision tasks. However, many open questions exist, including how to best utilize and improve the methods. Since similarity is ambiguous and context-dependent, it is also uncertain whether the methods can handle changing contexts.

This thesis addresses these questions through (i) a systematic study of different implementations of deep perceptual loss and similarity, (ii) a qualitative analysis of the strengths and weaknesses of the methods, (iii) a proof-of-concept investigation of the method's ability to adapt to new contexts, and (iv) cross-referencing the findings with already published works.

Several interesting findings are presented and discussed, including those below. Deep perceptual loss and similarity are shown not to follow existing transfer learning conventions. Flaws of the methods are discovered and mitigated. Deep perceptual similarity is demonstrated to be well-suited for applications in various contexts.

There is much left to explore, and this thesis provides insight into what future research directions are promising. Many improvements to deep perceptual similarity remain to be applied to loss calculation. Studying how related fields have dealt with problems caused by ambiguity and contexts could lead to further improvements. Combining these improvements could lead to metrics that perform close to the optimum on existing datasets, which motivates the development of more challenging datasets.

Measuring the similarity of images is a fundamental problem to computer vision for which no universal solution exists. While simple metrics such as the pixel-wise L2-norm have been shown to have significant flaws, they remain popular. One group of recent state-of-the-art metrics that mitigates some of those flaws are Deep Perceptual Similarity (DPS) metrics, where the similarity is evaluated as the distance in the deep features of neural networks.However, DPS metrics themselves have been less thoroughly examined for their benefits and, especially, their flaws. This work investigates the most common DPS metric, where deep features are compared by spatial position, along with metrics comparing the averaged and sorted deep features. The metrics are analyzed in-depth to understand the strengths and weaknesses of the metrics by using images designed specifically to challenge them. This work contributes with new insights into the flaws of DPS, and further suggests improvements to the metrics. An implementation of this work is available online: https://github.com/guspih/deep_perceptual_similarity_analysis/

This work presents a novel self-supervised pre-training method to learn efficient representations without labels on histopathology medical images utilizing magnification factors. Other state-of-the-art works mainly focus on fully supervised learning approaches that rely heavily on human annotations. However, the scarcity of labeled and unlabeled data is a long-standing challenge in histopathology. Currently, representation learning without labels remains unexplored in the histopathology domain. The proposed method, Magnification Prior Contrastive Similarity (MPCS), enables self-supervised learning of representations without labels on small-scale breast cancer dataset BreakHis by exploiting magnification factor, inductive transfer, and reducing human prior. The proposed method matches fully supervised learning state-of-the-art performance in malignancy classification when only 20% of labels are used in fine-tuning and outperform previous works in fully supervised learning settings for three public breast cancer datasets, including BreakHis. Further, It provides initial support for a hypothesis that reducing human-prior leads to efficient representation learning in self-supervision, which will need further investigation. The implementation of this work is available online on GitHub

Automatically identifying harmful content in video is an important task with a wide range of applications. However, there is a lack of professionally labeled open datasets available. In this work VidHarm, an open dataset of 3589 video clips from film trailers annotated by professionals, is presented. An analysis of the dataset is performed, revealing among other things the relation between clip and trailer level annotations. Audiovisual models are trained on the dataset and an in-depth study of modeling choices conducted. The results show that performance is greatly improved by combining the visual and audio modality, pre-training on large-scale video recognition datasets, and class balanced sampling. Lastly, biases of the trained models are investigated using discrimination probing.VidHarm is openly available, and further details are available at the webpage https://vidharm.github.io/

This work investigates three methods for calculating loss for autoencoder-based pretraining of image encoders: The commonly used reconstruction loss, the more recently introduced deep perceptual similarity loss, and a feature prediction loss proposed here; the latter turning out to be the most efficient choice. Standard auto-encoder pretraining for deep learning tasks is done by comparing the input image and the reconstructed image. Recent work shows that predictions based on embeddings generated by image autoencoders can be improved by training with perceptual loss, i.e., by adding a loss network after the decoding step. So far the autoencoders trained with loss networks implemented an explicit comparison of the original and reconstructed images using the loss network. However, given such a loss network we show that there is no need for the time-consuming task of decoding the entire image. Instead, we propose to decode the features of the loss network, hence the name “feature prediction loss”. To evaluate this method we perform experiments on three standard publicly available datasets (LunarLander-v2, STL-10, and SVHN) and compare six different procedures for training image encoders (pixel-wise, perceptual similarity, and feature prediction losses; combined with two variations of image and feature encoding/decoding). The embedding-based prediction results show that encoders trained with feature prediction loss is as good or better than those trained with the other two losses. Additionally, the encoder is significantly faster to train using feature prediction loss in comparison to the other losses. The method implementation used in this work is available online. https://github.com/guspih/Perceptual-Autoencoders

Autoencoders are commonly trained using element-wise loss. However, element-wise loss disregards high-level structures in the image which can lead to embeddings that disregard them as well. A recent improvement to autoencoders that helps alleviate this problem is the use of perceptual loss. This work investigates perceptual loss from the perspective of encoder embeddings themselves. Autoencoders are trained to embed images from three different computer vision datasets using perceptual loss based on a pretrained model as well as pixel-wise loss. A host of different predictors are trained to perform object positioning and classification on the datasets given the embedded images as input. The two kinds of losses are evaluated by comparing how the predictors performed with embeddings from the differently trained autoencoders. The results show that, in the image domain, the embeddings generated by autoencoders trained with perceptual loss enable more accurate predictions than those trained with element-wise loss. Furthermore, the results show that, on the task of object positioning of a small-scale feature, perceptual loss can improve the results by a factor 10. The experimental setup is available online: https://github.com/guspih/Perceptual-Autoencoders

In this paper, we introduce the use of Semantic Hashing as embedding for the task of Intent Classification and achieve state-of-the-art performance on three frequently used benchmarks. Intent Classification on a small dataset is a challenging task for data-hungry state-of-the-art Deep Learning based systems. Semantic Hashing is an attempt to overcome such a challenge and learn robust text classification. Current word embedding based methods [11], [13], [14] are dependent on vocabularies. One of the major drawbacks of such methods is out-of-vocabulary terms, especially when having small training datasets and using a wider vocabulary. This is the case in Intent Classification for chatbots, where typically small datasets are extracted from internet communication. Two problems arise with the use of internet communication. First, such datasets miss a lot of terms in the vocabulary to use word embeddings efficiently. Second, users frequently make spelling errors. Typically, the models for intent classification are not trained with spelling errors and it is difficult to think about ways in which users will make mistakes. Models depending on a word vocabulary will always face such issues. An ideal classifier should handle spelling errors inherently. With Semantic Hashing, we overcome these challenges and achieve state-of-the-art results on three datasets: Chatbot, Ask Ubuntu, and Web Applications [3]. Our benchmarks are available online.