Content-aware convolutional neural networks

Convolutional Neural Networks (CNNs) have achieved great success due to the powerful feature learning ability of convolution layers. Specifically, standard traverses input images/features using a sliding window scheme extract features. However, not all windows contribute equally prediction results CNNs. In practice, convolutional operation on some (e.g., smooth that contain very similar pixels) can be redundant and may introduce noises into computation. Such redundancy only deteriorate performance but also incur unnecessary computational cost. Thus, it is important reduce improve performance. To this end, we propose Content-aware Convolution (CAC) automatically detects applies 1 ×1 kernel replace original large kernel. sense, are able effectively avoid computation pixels. By replacing in CNNs with our CAC, resultant models yield significantly better lower cost than baseline convolution. More critically, dynamically allocate suitable resources according data smoothness different images, making possible for content-aware Extensive experiments various computer vision tasks demonstrate superiority method over existing methods.