Essential element-region driven model in image recognition

Abstract

The performance of image recognition depends on not only algorithms but also image factors, especially the essential elements and regions. This paper presents an essential element-region (EssER) recognition method. Firstly, we propose a series of contrast experiments to estimate the differences of the elements which include part, structure, texture and optical field. And the result shows that the key parts are influential in recognition tasks. Then for each input image, we apply a backtracking approach modeled by Convolutional Neural Networks (CNN) to calculate the essential regions in images. When only feeding the essential region instead of the original image into convolutional networks as input, the recognition result shows the performance decreases limitedly. At last, we propose a knowledge-guided attention network (KGAN) based on the EssER model to solve the problem of fine-grained classification and compared to traditional methods. Our model improves the performance at 4.7%, 6.0% than original baseline neural network with close computational cost on Stanford Cars, CUB-2011. Our method can reach the state-of-the-art performance in fine-grained classification.