Search

We propose an efficient neural network for RAW image denoising. Although neural network-based denoising has been extensively studied for image restoration, little attention has been given to efficient denoising for compute limited and power sensitive devices, such as smartphones and smartwatches. In this paper, we present a novel architecture and a suite of training techniques for high quality denoising in mobile devices. Our work is distinguished by three main contributions. (1) Feature-Align layer that modulates the activations of an encoder-decoder architecture with the input noisy images. The auto modulation layer enforces attention to spatially varying noise that tend to be "washed away" by successive application of convolutions and non-linearity. (2) A novel Feature Matching Loss that allows knowledge distillation from large denoising networks in the form of a perceptual content loss. (3) Empirical analysis of our efficient model trained to specialize on different noise subranges. This opens additional avenue for model size reduction by sacrificing memory for compute. Extensive experimental validation shows that our efficient model produces high quality denoising results that compete with state-of-the-art large networks, while using significantly fewer parameters and MACs. On the Darmstadt Noise Dataset benchmark, we achieve a PSNR of 48.28dB, while using 263 times fewer MACs, and 17.6 times fewer parameters than the state-of-the-art network, which achieves 49.12dB.

Adult colour preference has been summarized quantitatively in terms of weights on the two fundamental neural processes that underlie early colour encoding: the S-(L+M) ('blue-yellow') and L-M ('red-green') cone-opponent contrast channels (Ling, Hurlbert & Robinson, 2006; Hurlbert & Ling, 2007). Here, we investigate whether colour preference in 4-5-month-olds may be analysed in the same way. We recorded infants' eye-movements in response to pairwise presentations of eight colour stimuli varying only in hue. Infants looked longest at reddish and shortest at greenish hues. Analyses revealed that the L-M and S-(L+M) contrast between stimulus colour and background explained around half of the variation in infant preference across the hue spectrum. Unlike adult colour preference patterns, there was no evidence for sex differences in the weights on either of the cone-opponent contrast components. The findings provide a quantitative model of infant colour preference that summarizes variation in infant preference across hues.

Summary The long history of color preference studies has been described as "bewildering, confused and contradictory" [1]. Although recent studies [1–3] tend to agree on a universal preference for ‘blue', the variety and lack of control in measurement methods have made it difficult to extract a systematic, quantitative description of preference. Furthermore, despite abundant evidence for sex differences in other visual domains, and specifically in other tasks of color perception [4,5], there is no conclusive evidence for the existence of sex differences in color preference. This fact is perhaps surprising, given the prevalence and longevity of the notion that little girls differ from boys in preferring ‘pink' [6]. Here we report a robust, cross-cultural sex difference in color preference, revealed by a rapid paired-comparison task. Individual color preference patterns are summarized by weights on the two fundamental neural dimensions that underlie color coding in the human visual system. We find a consistent sex difference in these weights, which, we suggest, may be linked to the evolution of sex-specific behavioral uses of trichromacy.

We investigate color constancy for real 2D paper samples using a successive matching paradigm in which the observer memorizes a reference surface color under neutral illumination and after a temporal interval selects a matching test surface under the same or different illumination. We find significant effects of the illumination, reference surface, and their interaction on the matching error. We characterize the matching error in the absence of illumination change as the "pure color memory shift" and introduce a new index for successive color constancy that compares this shift against the matching error under changing illumination. The index also incorporates the vector direction of the matching errors in chromaticity space, unlike the traditional constancy index. With this index, we find that color constancy is nearly perfect.

In the natural world, objects are characterized by a variety of attributes, including color and shape. The contributions of these two attributes to object recognition are typically studied independently of each other, yet they are likely to interact in natural tasks. Here we examine whether color and size (a component of shape) interact in a real three-dimensional (3D) object similarity task, using solid domelike objects whose distinct apparent surface colors are independently controlled via spatially restricted illumination from a data projector hidden to the observer. The novel experimental setup preserves natural cues to 3D shape from shading, binocular disparity, motion parallax, and surface texture cues, while also providing the flexibility and ease of computer control. Observers performed three distinct tasks: two unimodal discrimination tasks, and an object similarity task. Depending on the task, the observer was instructed to select the indicated alternative object which was "bigger than," "the same color as," or "most similar to" the designated reference object, all of which varied in both size and color between trials. For both unimodal discrimination tasks, discrimination thresholds for the tested attribute (e.g., color) were increased by differences in the secondary attribute (e.g., size), although this effect was more robust in the color task. For the unimodal size-discrimination task, the strongest effects of the secondary attribute (color) occurred as a perceptual bias, which we call the "saturation-size effect": Objects with more saturated colors appear larger than objects with less saturated colors. In the object similarity task, discrimination thresholds for color or size differences were significantly larger than in the unimodal discrimination tasks. We conclude that color and size interact in determining object similarity, and are effectively analyzed on a coarser scale, due to noise in the similarity estimates of the individual attributes, inter-attribute attentional interactions, or coarser coding of attributes at a "higher" level of object representation.