Random Interpolation Resize: A free image data augmentation method for object detection in industry

Abstract

In the inspection task of industrial machine vision, insufficient data frequently occurs, and data enhancement can artificially incorporate the prior knowledge of human vision, expand the image data, and improve the model’s performance, which has become the model’s standard configuration. However, the majority of current data enhancement methods are intended for general scenes, and there are few image enhancement methods designed particularly for industrial object detection. Furthermore, many data enhancement methods require additional training time or annotation recalculation. This paper proposes a novel image data enhancement method, Random Interpolation Resize (RIR). We alter the interpolation method of the standard resize step in the preprocessing from fixed collocation to random selection in order to expand the image data and improve the model’s generalization and detection capabilities. There are no online editing options for annotations, and no additional training time is available. Multiple data sets were tested using various target detection algorithms, resulting in a 0.45% to 5.6% enhancement in mAP0.5 and an increase in the detection effect.