Title:Adaptive Neural Layer for Global Filtering in Computer Vision

Abstract:This study is motivated by typical images taken during ultrasonic examinations in the clinic. Their grainy appearance, low resolution, and poor contrast demand an eye of a very qualified expert to discern patterns and to spot pathologies. Automating any computer vision task on such data always involves excessive pre-processing and batch filtering, with an accumulation of the error emerging due to the annotation uncertainty, the digital post-filtering artifacts, and the amplified noise. Each patient case generally requires an individually tuned frequency filter to obtain optimal image contrast and to yield the desired outcome in a given computer vision problem. Thus, we aspired to invent a universal adaptive neural layer to "learn" optimal frequency filter for each image together with the weights of the network itself. The proposed approach takes the source image in the spatial domain, automatically selects the necessary frequencies from the frequency domain, and transmits the inverse-transform image to the main convolutional neural network. Remarkably, such a simple add-on layer can dramatically improve the performance of the main neural network regardless of its architecture. We observe that the light networks gain a noticeable boost in the performance metrics; whereas, the training of the heavy ones starts to converge faster when our adaptive layer is allowed to 'learn' alongside the main architecture. We validate the idea in four classical computer vision tasks: classification, segmentation, denoising, and erasing, considering popular natural and medical datasets.