Measuring the contribution of filter bank layer to performance of convolutional neural networks

Object identification is essential in diverse automated applications such as in health, business, and national security. It relies on the ability of the image processing scheme to detect visual features under a wide variety of conditions such as the object rotation, translation and geometric transformation. Machine learning methods, in this case, play an important role in improving the object identification performance by resolving whether the extracted visual patterns are from the possibly distorted target object or not. In recent works, systems that employ a Convolutional Neural Network (CNN) as the primary pattern recognition scheme demonstrate superior performance over other object identification systems based on handpicked shape-based features. Several studies credit this to the invariance of CNN to small distortion and spatial translation which in turn is attributed to its filter bank layer or the convolution layer. However, there has been no study to carefully test this claim. Towards studying the source of CNN\u27s superior performance, a methodology is designed that tracks the CNN performance when spatial information for visual features (e.g. edges, corners and end points) are gradually removed. Using the MNIST dataset, results show that as the spatial correlation information among pixels is slowly decreased, the performance of the CNN in recognizing handwritten digits also correspondingly decreases. The drop in accuracy continues until the accuracy approximates the performance of the classifier that was obtained without the filter bank. Conducted using a more complex dataset consisting of images of land vehicles, a similar set of experiments show the same drop in classification performance as spatial information among pixels is slowly removed. © 2017 - IOS Press and the authors. All rights reserved