Subjective image quality measures based on deep neural networks are very related to models of visual neuroscience. This connection benefits engineering but, more interestingly, the freedom to optimize deep networks in different ways, make them an excellent tool to explore the principles behind visual perception (both human and artificial). Recently, a myriad of networks have been successfully optimized for many interesting visual tasks. Although these nets were not specifically designed to predict image quality or other psychophysics, they have shown surprising human-like behavior. The reasons for this remain unclear. In this work, we perform a thorough analysis of the perceptual properties of pre-trained nets (particularly their ability to predict image quality) by isolating different factors: the goal (the function), the data (learning environment), the architecture, and the readout: selected layer(s), fine-tuning of channel relevance, and use of statistical descriptors as opposed to plain readout of responses. Several conclusions can be drawn. All the models correlate better with human opinion than SSIM. More importantly, some of the nets are in pair of state-of-the-art with no extra refinement or perceptual information. Nets trained for supervised tasks such as classification correlate substantially better with humans than LPIPS (a net specifically tuned for image quality). Interestingly, self-supervised tasks such as jigsaw also perform better than LPIPS. Simpler architectures are better than very deep nets. In simpler nets, correlation with humans increases with depth as if deeper layers were closer to human judgement. This is not true in very deep nets. Consistently with reports on illusions and contrast sensitivity, small changes in the image environment does not make a big difference. Finally, the explored statistical descriptors and concatenations had no major impact.
High-resolution (HR) images are usually downscaled to low-resolution (LR) ones for better display and afterward upscaled back to the original size to recover details. Recent work in image rescaling formulates downscaling and upscaling as a unified task and learns a bijective mapping between HR and LR via invertible networks. However, in real-world applications (e.g., social media), most images are compressed for transmission. Lossy compression will lead to irreversible information loss on LR images, hence damaging the inverse upscaling procedure and degrading the reconstruction accuracy. In this paper, we propose the Self-Asymmetric Invertible Network (SAIN) for compression-aware image rescaling. To tackle the distribution shift, we first develop an end-to-end asymmetric framework with two separate bijective mappings for high-quality and compressed LR images, respectively. Then, based on empirical analysis of this framework, we model the distribution of the lost information (including downscaling and compression) using isotropic Gaussian mixtures and propose the Enhanced Invertible Block to derive high-quality/compressed LR images in one forward pass. Besides, we design a set of losses to regularize the learned LR images and enhance the invertibility. Extensive experiments demonstrate the consistent improvements of SAIN across various image rescaling datasets in terms of both quantitative and qualitative evaluation under standard image compression formats (i.e., JPEG and WebP).
Ultrasound images are widespread in medical diagnosis for musculoskeletal, cardiac, and obstetrical imaging due to the efficiency and non-invasiveness of the acquisition methodology. However, the acquired images are degraded by acoustic (e.g. reverberation and clutter) and electronic sources of noise. To improve the Peak Signal to Noise Ratio (PSNR) of the images, previous denoising methods often remove the speckles, which could be informative for radiologists and also for quantitative ultrasound. Herein, a method based on the recent Denoising Diffusion Probabilistic Models (DDPM) is proposed. It iteratively enhances the image quality by eliminating the noise while preserving the speckle texture. It is worth noting that the proposed method is trained in a completely unsupervised manner, and no annotated data is required. The experimental blind test results show that our method outperforms the previous nonlocal means denoising methods in terms of PSNR and Generalized Contrast to Noise Ratio (GCNR) while preserving speckles.
Federated learning (FL) enables multiple sites to collaboratively train powerful deep models without compromising data privacy and security. The statistical heterogeneity (e.g., non-IID data and domain shifts) is a primary obstacle in FL, impairing the generalization performance of the global model. Weakly supervised segmentation, which uses sparsely-grained (i.e., point-, bounding box-, scribble-, block-wise) supervision, is increasingly being paid attention to due to its great potential of reducing annotation costs. However, there may exist label heterogeneity, i.e., different annotation forms across sites. In this paper, we propose a novel personalized FL framework for medical image segmentation, named FedICRA, which uniformly leverages heterogeneous weak supervision via adaptIve Contrastive Representation and Aggregation. Concretely, to facilitate personalized modeling and to avoid confusion, a channel selection based site contrastive representation module is employed to adaptively cluster intra-site embeddings and separate inter-site ones. To effectively integrate the common knowledge from the global model with the unique knowledge from each local model, an adaptive aggregation module is applied for updating and initializing local models at the element level. Additionally, a weakly supervised objective function that leverages a multiscale tree energy loss and a gated CRF loss is employed to generate more precise pseudo-labels and further boost the segmentation performance. Through extensive experiments on two distinct medical image segmentation tasks of different modalities, the proposed FedICRA demonstrates overwhelming performance over other state-of-the-art personalized FL methods. Its performance even approaches that of fully supervised training on centralized data. Our code and data are available at https://github.com/llmir/FedICRA.
Recently, large-scale vision-language pre-training models and visual semantic embedding methods have significantly improved image-text matching (ITM) accuracy on MS COCO 5K test set. However, it is unclear how robust these state-of-the-art (SOTA) models are when using them in the wild. In this paper, we propose a novel evaluation benchmark to stress-test the robustness of ITM models. To this end, we add various fooling images and captions to a retrieval pool. Specifically, we change images by inserting unrelated images, and change captions by substituting a noun, which can change the meaning of a sentence. We discover that just adding these newly created images and captions to the test set can degrade performances (i.e., Recall@1) of a wide range of SOTA models (e.g., 81.9% $\rightarrow$ 64.5% in BLIP, 66.1% $\rightarrow$ 37.5% in VSE$\infty$). We expect that our findings can provide insights for improving the robustness of the vision-language models and devising more diverse stress-test methods in cross-modal retrieval task. Source code and dataset will be available at https://github.com/pseulki/rococo.
Training segmentation models for medical images continues to be challenging due to the limited availability and acquisition expense of data annotations. Segment Anything Model (SAM) is a foundation model trained on over 1 billion annotations, predominantly for natural images, that is intended to be able to segment the user-defined object of interest in an interactive manner. Despite its impressive performance on natural images, it is unclear how the model is affected when shifting to medical image domains. Here, we perform an extensive evaluation of SAM's ability to segment medical images on a collection of 11 medical imaging datasets from various modalities and anatomies. In our experiments, we generated point prompts using a standard method that simulates interactive segmentation. Experimental results show that SAM's performance based on single prompts highly varies depending on the task and the dataset, i.e., from 0.1135 for a spine MRI dataset to 0.8650 for a hip x-ray dataset, evaluated by IoU. Performance appears to be high for tasks including well-circumscribed objects with unambiguous prompts and poorer in many other scenarios such as segmentation of tumors. When multiple prompts are provided, performance improves only slightly overall, but more so for datasets where the object is not contiguous. An additional comparison to RITM showed a much better performance of SAM for one prompt but a similar performance of the two methods for a larger number of prompts. We conclude that SAM shows impressive performance for some datasets given the zero-shot learning setup but poor to moderate performance for multiple other datasets. While SAM as a model and as a learning paradigm might be impactful in the medical imaging domain, extensive research is needed to identify the proper ways of adapting it in this domain.
Whereas machine learning models typically learn language by directly training on language tasks (e.g., next-word prediction), language emerges in human children as a byproduct of solving non-language tasks (e.g., acquiring food). Motivated by this observation, we ask: can embodied reinforcement learning (RL) agents also indirectly learn language from non-language tasks? Learning to associate language with its meaning requires a dynamic environment with varied language. Therefore, we investigate this question in a multi-task environment with language that varies across the different tasks. Specifically, we design an office navigation environment, where the agent's goal is to find a particular office, and office locations differ in different buildings (i.e., tasks). Each building includes a floor plan with a simple language description of the goal office's location, which can be visually read as an RGB image when visited. We find RL agents indeed are able to indirectly learn language. Agents trained with current meta-RL algorithms successfully generalize to reading floor plans with held-out layouts and language phrases, and quickly navigate to the correct office, despite receiving no direct language supervision.
Recent camera-based 3D object detection methods have introduced sequential frames to improve the detection performance hoping that multiple frames would mitigate the large depth estimation error. Despite improved detection performance, prior works rely on naive fusion methods (e.g., concatenation) or are limited to static scenes (e.g., temporal stereo), neglecting the importance of the motion cue of objects. These approaches do not fully exploit the potential of sequential images and show limited performance improvements. To address this limitation, we propose a novel 3D object detection model, P2D (Predict to Detect), that integrates a prediction scheme into a detection framework to explicitly extract and leverage motion features. P2D predicts object information in the current frame using solely past frames to learn temporal motion features. We then introduce a novel temporal feature aggregation method that attentively exploits Bird's-Eye-View (BEV) features based on predicted object information, resulting in accurate 3D object detection. Experimental results demonstrate that P2D improves mAP and NDS by 3.0% and 3.7% compared to the sequential image-based baseline, illustrating that incorporating a prediction scheme can significantly improve detection accuracy.
Efficient deep learning-based approaches have achieved remarkable performance in single image super-resolution. However, recent studies on efficient super-resolution have mainly focused on reducing the number of parameters and floating-point operations through various network designs. Although these methods can decrease the number of parameters and floating-point operations, they may not necessarily reduce actual running time. To address this issue, we propose a novel multi-stage lightweight network boosting method, which can enable lightweight networks to achieve outstanding performance. Specifically, we leverage enhanced high-resolution output as additional supervision to improve the learning ability of lightweight student networks. Upon convergence of the student network, we further simplify our network structure to a more lightweight level using reparameterization techniques and iterative network pruning. Meanwhile, we adopt an effective lightweight network training strategy that combines multi-anchor distillation and progressive learning, enabling the lightweight network to achieve outstanding performance. Ultimately, our proposed method achieves the fastest inference time among all participants in the NTIRE 2023 efficient super-resolution challenge while maintaining competitive super-resolution performance. Additionally, extensive experiments are conducted to demonstrate the effectiveness of the proposed components. The results show that our approach achieves comparable performance in representative dataset DIV2K, both qualitatively and quantitatively, with faster inference and fewer number of network parameters.
Adversarial attacks on deep-learning models pose a serious threat to their reliability and security. Existing defense mechanisms are narrow addressing a specific type of attack or being vulnerable to sophisticated attacks. We propose a new defense mechanism that, while being focused on image-based classifiers, is general with respect to the cited category. It is rooted on hyperspace projection. In particular, our solution provides a pseudo-random projection of the original dataset into a new dataset. The proposed defense mechanism creates a set of diverse projected datasets, where each projected dataset is used to train a specific classifier, resulting in different trained classifiers with different decision boundaries. During testing, it randomly selects a classifier to test the input. Our approach does not sacrifice accuracy over legitimate input. Other than detailing and providing a thorough characterization of our defense mechanism, we also provide a proof of concept of using four optimization-based adversarial attacks (PGD, FGSM, IGSM, and C\&W) and a generative adversarial attack testing them on the MNIST dataset. Our experimental results show that our solution increases the robustness of deep learning models against adversarial attacks and significantly reduces the attack success rate by at least 89% for optimization attacks and 78% for generative attacks. We also analyze the relationship between the number of used hyperspaces and the efficacy of the defense mechanism. As expected, the two are positively correlated, offering an easy-to-tune parameter to enforce the desired level of security. The generality and scalability of our solution and adaptability to different attack scenarios, combined with the excellent achieved results, other than providing a robust defense against adversarial attacks on deep learning networks, also lay the groundwork for future research in the field.