Insect recognition is a hard problem because the difference of appearance between insects is so small that only some entomologist experts can distinguish them. Besides that, insects are often composed of several parts (multiple views) which generate more degrees of freedom. This chapter proposes several discriminative coding approaches and one decision fusion scheme of heterogeneous class sets for insect recognition. The three discriminative coding methods use class specific concatenated vectors instead of traditional global coding vectors for insect image patches. The decision fusion scheme uses an allocation matrix for classifier selection and a weight matrix for classifier fusion, which is suitable for combining classifiers of heterogeneous class sets in multi-view insect image recognition. Experimental results on a Tephritidae dataset show that the three proposed discriminative coding methods perform well in insect recognition, and the proposed fusion scheme improves the recognition accuracy significantly.
TopIntroduction
As the largest biological population all over the world, insects have close relationship to human life. Some kinds can take advantages to agriculture and ecology, while others may cause huge losses to them. Therefore a large scale insect surveillance system is essential to monitoring the environment. Because of the complex and confusing appearance anatomy, insect recognition by human experts needs not only professional knowledge, but also heavy labor. Therefore, automated insect recognition by computer vision techniques instead of human experts is more and more urgent in application. The goal of this research is to develop a computer vision system which is convenient (without much interaction such as alignment, rotation and segmentation), foolproof (not necessary to have expert knowledge), and inexpensive (only needing a PC, a digital camera and little human labor) to partly take place of human experts in some area and help entomologists lighten heavy labor in their researches.
Besides its practical importance, automated insect recognition also raises many fundamental computer vision challenges. Insects in the same or similar species have very alike appearance so that a laymen can not identify them. Invariance to pose, size, and orientation is also another problem to be solved in insect recognition and many other computer vision applications. These lead to the challenge of large intra-category variations and small inter-category differences (Larios, et al., 2008). The basic construction methods of this chapter are motivated by sparse representation based classification (SRC) algorithm for face recognition by Wright (Wright, Yang, Ganesh, Sastry, & Ma, 2009) and image restoration method by Mairal (Mairal, Elad, & Sapiro, 2008). Both of the two methods are based on sparse representation which is very popular in computer vision, image processing and machine learning. The term “sparse representation” refers to an expression of the input signal as a linear combination of basis elements in which many of the coefficients are zero (Wright, et al., 2009). Different to (Wright, et al., 2009) and (Mairal, et al., 2008), the authors of this chapter calculate bases of each class to obtain the class specific representations in the strategy of minimal reconstruction residual of local features combined with sparse coding by Yang (Yang, Yu, Gong, & Huang, 2009), soft coding by Gemert (Gemert, Geusebroek, Veenman, & Smeulders, 2008), and salient coding by Huang (Huang, Huang, Yu, & Tan, 2011). All of the three methods sparse coding, soft coding and salient coding lose some discriminative information in the codebook generation process and local information in the subsequent coding process. So the authors make use of class specific codebooks to improve the discriminability of codewords and get more robust representations of feature vectors with smaller reconstruction error.