Anti-aliasing deep image classifiers using novel depth adaptive blurring and activation function

Deep convolutional networks are vulnerable to image translation or shift, partly due common down-sampling layers, e.g., max-pooling and strided convolution. These operations violate the Nyquist sampling rate cause aliasing. The textbook solution is low-pass filtering (blurring) before down-sampling, which can benefit deep as well. Even so, non-linearity units, such ReLU, often re-introduce problem, suggesting that blurring alone may not suffice. In this work, first, we analyse features with Fourier transform show Depth Adaptive Blurring more effective, opposed monotonic blurring. To end, propose a novel Blur-pool (DAB-pool) module replace existing methods. Second, introduce activation function – built-in low pass filter, an additional measure, keep problem from reappearing. From experiments, observe generalisation on other forms of transformations corruptions well, rotation, scale, noise. We evaluate our method under three challenging settings: (1) variety translations; (2) adversarial attacks both ℓp bounded unbounded; (3) data perturbations. each setting, achieves state-of-the-art results improves clean accuracy various benchmark datasets.