Dirty Pixels: Optimizing Image Classification Architectures for Raw Sensor Data

By Steven Diamond, Vincent Sitzmann, Stephen P. Boyd, Gordon Wetzstein, Felix Heide. 2017

95 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

A Study and Comparison of Human and Deep Learning Recognition Performance under Visual Distortions (2017, background): “Finally, the dirty pixel approach [16] shows that ﬁne-tuning with an additional pre-processing module can yield a DNN more robust to blur and noise.”
Google's Cloud Vision API is Not Robust to Noise (2017, methodology): “For improving the robustness of machine learning models to input perturbations, an end-to-end architecture is proposed in [16] for joint denoising, deblurring, and classiﬁcation.”
AOD-Net: All-in-One Dehazing Network (2017, background): “Recently, [27, 4] validated that a joint optimization of restoration and recognition steps would boost the performance over the traditional two-stage approach.”
End-to-end optimization of optics and image processing for achromatic extended depth of field and super-resolution imaging (2018, methodology): “2017b] could be used in this model or image reconstructions could potentially be replaced with a high-level convolutional neural network for image analysis [Diamond et al. 2017a].”
Deflecting Adversarial Attacks with Pixel Deflection (2018, background): “However, most deep classifiers are robust to the presence of natural noise, such as sensor noise [10].”
Neural Nearest Neighbors Networks (2018, background): “Besides improving the visual quality of noisy images, the importance of image denoising also stems from the fact that image noise severely degrades the accuracy of downstream computer vision tasks, e. g. detection [10].”
DesnowNet: Context-Aware Deep Network for Snow Removal (2017, methodology): “To this end, we first employ Inception-v4 [28] as the backbone due to its highly optimized features at multi-scale receptive fields, a network whose success has been demonstrated in diverse studies [2, 10] To further extend context-awareness and exploit multi-scale features, we proposed a new subnet”
Human and DNN Classification Performance on Images With Quality Distortions (2019, methodology): “Another approach is to perform some preprocessing to transform the distorted image into a cleaner representation to input to a DNN (Diamond et al. 2016).”
Quality Resilient Deep Neural Networks (2017, background): “[2] describe a system that prepends a neural network with additional layers that serve to undistort the image.”