Global-first Training Strategy with Convolutional Neural Networks to Improve Scale Invariance

Abstract

Modelled closely on the feedforward conical structure of the primate vision system - Convolutional Neural Networks (CNNs) learn by adopting a local to global feature extraction strategy. This makes them view-specific models and results in poor invariance encoding within its learnt weights to adequately identify objects whose appearance is altered by various transformations such as rotations, translations, and scale. Recent physiological studies reveal the visual system first views the scene globally for subsequent processing in its ventral stream leading to a global-first response strategy in its recognition function. Conventional CNNs generally use small filters, thus losing the global view of the image. A trainable module proposed by Kumar & Sharma [24] called Stacked Filters Convolution (SFC) models this approach by using a pyramid of large multi-scale filters to extract features from wider areas of the image, which is then trained by a normal CNN. The end-to-end model is referred to as Stacked Filter CNN (SFCNN). In addition to improved test results, SFCNN showed promising results on scale invariance classification. The experiments, however, were performed on small resolution datasets and small CNN as backbone. In this paper, we extend this work and test SFC integrated with the VGG16 network on larger resolution datasets for scale invariance classification. Our results confirm the integration of SFC, and standard CNN also shows promising results on scale invariance on large resolution datasets.