Rectify ViT Shortcut Learning by Visual Saliency.

Data storage via multimedia technology is more preferred as the information in multimedia contain rich meanings and are concise when compared to the traditional textual information. However, efficient information retrieval is a crucial factor in such storage. This paper presents a cognitive classification based visual saliency guided model for the efficient retrieval of information from multimedia data storage. The Itti visual saliency model is described here for generation of an overall saliency map with the integration of color saliency, intensity and direction maps. Multi-feature fusion paradigms are used for providing clear description of the image pattern. The definition is based on two stages namely complexity based on cognitive load and classification of complexity at a cognitive level. The image retrieval system is finalized by integrating a group sparse logistic regression model. In complex scenarios, the baselines are overcome by the proposed system when tested on multiple databased as compared to other state-of-the-art models.

Healthcare is seen as one of the most influential applications of Deep Learning (DL). Increasingly, DL models have been shown to achieve high-levels of performance on medical diagnosis tasks, in some cases achieving levels of performance on-par with medical experts. Yet, very few are deployed into real-life scenarios. One of the main reasons for this is the lack of trust in those models by medical professionals driven by the black-box nature of the deployed models. Numerous explainability techniques have been developed to alleviate this issue by providing a view on how the model reached a given decision. Recent studies have shown that those explanations can expose the models’ reliance on areas of the feature space that has no justifiable medical interpretation, widening the gap with the medical experts. In this paper we evaluate the deviation of saliency maps produced by DL classification models from radiologist’s eye-gaze while they study the MIMIC-CXR-EGD images, and we propose a novel model architecture that utilises model explanations during training only (i.e. not during inference) to improve the overall plausibility of the model explanations. We substantially improve the similarity between the model’s explanations and radiologists’ eye-gaze data, reducing Kullback-Leibler Divergence by 90% and increasing Normalised Scanpath Saliency by 216%. We argue that this significant improvement is an important step towards building more robust and interpretable DL solutions in health-care.

Modeling visual saliency map of an image provides important information for image semantic understanding in many applications. Most existing computational visual saliency models follow a bottom-up framework that generates independent saliency map in each selected visual feature space and combines them in a proper way. Two big challenges to be addressed explicitly in these methods are (1) which features should be extracted for all pixels of the input image and (2) how to dynamically determine importance of the saliency map generated in each feature space. In order to address these problems, we present a novel saliency map computational model based on tensor decomposition and reconstruction. Tensor representation and analysis not only explicitly represent image's color values but also imply two important relationships inherent to color image. One is reflecting spatial correlations between pixels and the other one is representing interplay between color channels. Therefore, saliency map generator based on the proposed model can adaptively find the most suitable features and their combinational coefficients for each pixel. Experiments on a synthetic image set and a real image set show that our method is superior or comparable to other prevailing saliency map models.

Visual saliency distribution maps for explaining time-series AI models used in continuous production of textile fibers

This article proposes a bottom-up visual saliency model that uses the wavelet transform to conduct multiscale analysis and computation in the frequency domain. First, we compute the multiscale magnitude spectra by performing a wavelet transform to decompose the magnitude spectrum of the discrete cosine coefficients of an input image. Next, we obtain multiple saliency maps of different spatial scales through an inverse transformation from the frequency domain to the spatial domain, which utilizes the discrete cosine magnitude spectra after multiscale wavelet decomposition. Then, we employ an evaluation function to automatically select the two best multiscale saliency maps. A final saliency map is generated via an adaptive integration of the two selected multiscale saliency maps. The proposed model is fast, efficient, and can simultaneously detect salient regions or objects of different sizes. It outperforms state-of-the-art bottom-up saliency approaches in the experiments of psychophysical consistency, eye fixation prediction, and saliency detection for natural images. In addition, the proposed model is applied to automatic ship detection in optical satellite images. Ship detection tests on satellite data of visual optical spectrum not only demonstrate our saliency model's effectiveness in detecting small and large salient targets but also verify its robustness against various sea background disturbances.

Image saliency estimation via random walk guided by informativeness and latent signal correlations

Characteristics of the human visual field are well known to be different in central (fovea) and peripheral areas. Existing computational models of visual saliency, however, do not take into account this biological evidence. The existing models compute visual saliency uniformly over the retina and, thus, have difficulty in accurately predicting the next gaze (fixation) point. This paper proposes to incorporate human visual field characteristics into visual saliency, and presents a computational model for producing such a saliency map. Our model integrates image features obtained by bottom-up computation in such a way that weights for the integration depend on the distance from the current gaze point where the weights are optimally learned using actual saccade data. The experimental results using a large number of fixation/saccade data with wide viewing angles demonstrate the advantage of our saliency map, showing that it can accurately predict the point where one looks next.

Several computational visual saliency models have been proposed in the context of viewing natural scenes. We aim to investigate the relevance of computational saliency models in medical images in the context of abnormality detection. We report on two studies aimed at understanding the role of visual saliency in medical images. Diffuse lesions in Chest X-Ray images, which are characteristic of Pneumoconiosis and high contrast lesions such as 'Hard Exudates' in retinal images were chosen for the study. These approximately correspond to conjunctive and disjunctive targets in a visual search task. Saliency maps were computed using three popular models namely Itti-Koch [7], GBVS [3] and SR [4]. The obtained maps were evaluated against gaze maps and ground truth from medical experts.Our results show that GBVS is seen to perform the best (Mdn. ROC area = 0.77) for chest X-Ray images while SR performs the best (ROC area = 0.73) for retinal images, thus asserting that searching for conjunctive targets calls for a more local examination of an image while disjunctive targets call for a global examination. Based on the results of the above study, we propose extensions for the two best performing models. The first extension makes use of top down knowledge such as lung segmentation. This is shown to improve the performance of GBVS to some extent. In the second case the extension is by way of including multi-scale information. This is shown to significantly (by 28.76%) improve abnormality detection. The key insight from these studies is that bottom saliency continues to play a predominant role in examining medical images.

Visual saliency map has been proposed as a computational model for estimating the visual attention of a human. The saliency map is used to estimate the bottom-up visual attention for still and moving images. However, under the condition including the top-down visual attention, e.g. advertisement, the accuracy of visual attention estimated from the saliency map decreased. The character feature is considered as one of the factors that the deterioration of the saliency map accuracy is induced. In this study, we hypothesized that the features of characters have the saliency to induce the visual attention. Two types of experiments were performed to test this hypothesis. The still images inserted with both of characters (HIRAGANAs in Japanese or Thai strings) and simple symbols were used as visual stimuli in the experiments. The visual stimuli were presented to the subjects for a short period (2s) to exclude the effect of top-down attention. In result, the fixation probability of the character (both HIRAGANAs and Thais) region in the image was higher than that of the symbol region. The paired t-test provided the significant difference of the fixation ratio between HIRAGANAs and symbols (p < 0.001). The same is true for the paired t-test between Thais and symbols (p < 0.001). Thus, the present results indicate the visual saliency of characters.

Radiologists' eye movements during medical image interpretation reflect their perceptual-cognitive processes of diagnostic decisions. The eye movement data can be modeled to represent clinically relevant regions in a medical image and potentially integrated into an artificial intelligence (AI) system for automatic diagnosis in medical imaging. In this article, we first conduct a large-scale eye-tracking study involving 13 radiologists interpreting 191 chest X-ray (CXR) images, establishing a best-of-its-kind CXR visual saliency benchmark. We then perform analysis to quantify the reliability and clinical relevance of saliency maps (SMs) generated for CXR images. We develop CXR image saliency prediction method (CXRSalNet), a novel saliency prediction model that leverages radiologists' gaze information to optimize the use of unlabeled CXR images, enhancing training and mitigating data scarcity. We also demonstrate the application of our CXR saliency model in enhancing the performance of AI-powered diagnostic imaging systems.

Evaluating, explaining, and visualizing high-level concepts in generative models, such as variational autoencoders (VAEs), is challenging in part due to a lack of known prediction classes that are required to generate saliency maps in supervised learning. While saliency maps may help identify relevant features (e.g., pixels) in the input for classification tasks of deep neural networks, similar frameworks are understudied in unsupervised learning. Therefore, we introduce a new method of obtaining saliency maps for latent representations of known or novel high-level concepts, often called concept vectors in generative models. Concept scores, analogous to class scores in classification tasks, are defined as dot products between concept vectors and encoded input data, which can be readily used to compute the gradients. The resulting concept saliency maps are shown to highlight input features deemed important for high-level concepts. Our method is applied to the VAE's latent space of CelebA dataset in which known attributes such as "smiles" and "hats" are used to elucidate relevant facial features. Furthermore, our application to spatial transcriptomic (ST) data of a mouse olfactory bulb demonstrates the potential of latent representations of morphological layers and molecular features in advancing our understanding of complex biological systems. By extending the popular method of saliency maps to generative models, the proposed concept saliency maps help improve interpretability of latent variable models in deep learning. Codes to reproduce and to implement concept saliency maps: https://github.com/lenbrocki/concept-saliency-maps

Objective video quality assessment plays an important role in a variety of video processing applications such as video compression, transmission, visualization, and display. This paper proposes a no-reference video quality assessment model based on visual memory understanding. Inspired by the findings of neuroscience researchers, who argue there is a large overlap between the active human brain area when performing video quality assessment and saliency detection tasks, saliency maps are employed here to assist the quality assessment. To this end, we first generate the saliency maps using CLBP (Complete Local Binary Patterns) features of the residual frames. Then, a model of visual memory is created from the statistics of saliency maps. This is followed by learning the video quality from the visual memory, saliency, and frame features through a support vector regression pipeline. The experimental results on the state-of-the-art LIVE and SJTU video datasets indicate that the proposed no-reference video quality assessment algorithm is effective and performs statistically better than several other state-of-the-art approaches.

In a general scenario, while attending a scene containing multiple faces or looking towards a group photograph, our attention does not go equal towards all the faces. It means, we are naturally biased towards some faces. This biasness happens due to availability of dominant perceptual features in those faces. In visual saliency terminology it can be called as 'salient face'. Human's focus their gaze towards a face which carries the 'dominating look' in the crowd. This happens due to comparative saliency of the faces. Saliency of a face is determined by its feature dissimilarity with the surrounding faces. In this context there is a big role of human psychology and its cognitive science too. Therefore, enormous researches have been carried out towards modeling the computer vision system like human's vision. This paper proposed a graphical based bottom up approach to point up the salient face in the crowd or in an image having multiple faces. In this novel method, visual saliencies of faces have been calculated based on the intensity values, facial areas and their relative spatial distances. Experiment has been conducted on gray scale images. In order to verify this experiment, three level of validation has been done. In the first level, our results have been verified with the prepared ground truth. In the second level, intensity scores of proposed saliency maps have been cross verified with the saliency score. In the third level, saliency map is validated with some standard parameters. The results are found to be interesting and in some aspects saliency predictions are like human vision system. The evaluation made with the proposed approach shows moderately boost up results and hence, this idea can be useful in the future modeling of intelligent vision (robot vision) system.

Visual saliency estimation through label propagation is proposed. Since salient regions are generally distinctive from the background, visual saliency is estimated in an image by the similarity to salient and non‐salient seeds. First, salient and non‐salient seeds from the initial saliency map estimated by superpixel segmentation and three common priors are obtained. Then, labels of seeds are propagated to unlabelled superpixels based on label propagation to obtain the visual saliency map. Experimental results demonstrate that the proposed method is very effective in estimating salient regions in images compared with existing state‐of‐the‐art methods.

Visual saliency detection plays a significant role in the fields of computer vision. In this paper, we introduce a novel saliency detection method based on weighted linear multiple kernel learning (WLMKL) framework, which is able to adaptively combine different contrast measurements in a supervised manner. As most influential factor is contrast operation in bottom-up visual saliency, an average weighted corner-surround contrast (AWCSC) is first designed to measure local visual saliency. Combined with common-used center-surrounding contrast (CESC) and global contrast (GC), three types of contrast operations are fed into our WLMKL framework to produce the final saliency map. We show that the assigned weights for each contrast feature maps are always normalized in our WLMKL formulation. In addition, the proposed approach benefits from the advantages of the contribution of each individual contrast feature maps, yielding more robust and accurate saliency maps. We evaluated our method for two main visual saliency detection tasks: human fixed eye prediction and salient object detection. The extensive experimental results show the effectiveness of the proposed model, and demonstrate the integration is superior than individual subcomponent.