A novel biologically inspired ELM-based network for image recognition

Abstract

In this paper, a novel biologically inspired network for image recognition has been introduced. The Hierarchical model and X (HMAX) model and the extreme learning machine (ELM) are combined, to construct a five-layer feed-forward network: S1–C1–S2–C2–H. The previous four layers, originating from HMAX, provide robust feature representation of specific object, and the feature classification stage in the H layer is implemented with ELM. The HMAX model simulates the hierarchical processing mechanism in primate visual cortex, to calculate complex features representation. As a biological learning algorithm for generalized SLFNs, ELM learns much faster with good generalization performance, and performs well in classification applications. Four groups of experiments are performed on three datasets, and the results are compared with state-of-the-art techniques. Experimental results show that our proposed network has good performance with fast learning speed.

Introduction

Object recognition has been a popular area of intense research and is also a very challenging task in computer vision, while human vision with unique processing mechanism has the ability to recognize objects rapidly, accurately, and effortlessly. The difficulty of object recognition in images is due to different illuminations, viewpoints, occlusions, scale and shift transforms. Meanwhile, the difficulty of object categorization lies in capturing the variability of appearance and shape of different objects belonging to the same class, while avoiding confusing objects from different classes. Thus in order to achieve robust object recognition, overcoming these obstacles above would be beneficial for many fields and applications, such as security surveillance, manufacturing production, robot navigation, character recognition, and clinical image understanding.

There are many research works done for object recognition. In Treiber׳s book [1], he gives a good overview of object recognition algorithms used in various applications, including global approaches, transformation-search-based methods, geometrical model driven methods, 3D object recognition schemes, flexible contour fitting algorithms, and descriptor-based methods. In the work of Belongie [2], a set of discrete points sampled from the contour of the shape is used as a shape descriptor and then K nearest neighbors (KNN) are used for classification. Mohan [3] built a parts based detector with Haar wavelets to represent the image, and then use a support vector machine (SVM) for classification. Lowe [4] developed an image feature, called scale-invariant feature transform (SIFT), that became the basis for features in many object recognition algorithms, while it is not an object recognition algorithm by itself. Laptev [5] uses histograms as features, weighted Fisher linear discriminant as a weak classifier, and then the AdaBoost for classification. Biologically motivated features based on Gabor filters and MAX operations have been developed [6], [7]. Although the performance of the object recognition has been improved with the above algorithms, none of these algorithms available today can surpass the performance of the human brain. It suggests that more work needs to be done in this field, to solve the multiple problems in robust intelligence that the human brain is so good at, for enhancing the performance of object recognition.

Object recognition in human brain is largely invariant with regard to changes in the size, position, and viewpoint of the object. Therefore, it is perhaps not too surprising that the human brain has achieved, through millions of years of evolution, a remarkable ability to recognize objects in a robust, selective and fast manner. It is likely that, upon understanding how the neuronal circuitries can achieve these remarkable properties, it will be possible to translate the biological circuits into algorithms for computer vision and pattern recognition. A hierarchical cortical based model, named Hierarchical Model and X (HMAX) [8], [9], has attracted much attention, as the fact that it focuses on designing simple and complex operations inspired by the visual cortex, and that in [10], it is shown that the HMAX can provide robust representations of specific images, outperforming state-of-the-art such as SIFT under various invariance tasks on synthetic images. Recently, a novel learning algorithm for single hidden layer feed-forward networks (SLFNs), namely, extreme learning machine (ELM), proposed by Huang et al. [11], can be applied to regression and classification problems [12]. And in [13], it has been successfully applied in the face recognition, which improves the recognition accuracy rate.

In order to improve the performance of image based object recognition, this paper brings together two biologically inspired algorithms, HMAX and ELM, and insights to construct a novel biologically inspired network for image recognition. Since the HMAX features have better scale and translation invariance, the four-layer HMAX model, is employed for feature construction, feature selection and feature extraction, and provides robust feature representation of specific object image. As it has better performance than conventional methods, such as SVM, and it has an extremely fast learning speed, which is akin to the fast learning mechanism of the higher cortical areas, ELM is introduced for feature representation classification. Four groups of experiments will be performed on three datasets, to demonstrate the novelty and superiority of our proposed network over existing algorithms.

The rest of the paper is organized as follows. Section 2 states the problem and strategy of object recognition. In Section 3, preliminary information about HMAX and ELM is presented. Section 4 details the proposed biologically inspired ELM-based image recognition algorithm. In Section 5, several experiments are performed, and followed by results and discussions. The paper is concluded in Section 6.