A Neural Network Incorporating Adaptive Gabor Filters for Image Texture Classification

the peak frequency [2]. When the textures are unstable, however, e.g, when peak frequencies have fluctuations, A novel neural network architecture for image simple comparison of the peak outputs may be ineffective. texture classification is introduced. The proposed Kernel The shortcomings of using the peak frequencies of each Modifying Neural Network (KM Net) which incorporates texture classes have been also discussed by Teuner et al., a convolution filter kernel array and a classifier in one, where the spectral features that are ìout of ordinaryî are enables an automated texture feature extraction in the extracted using the spectral feature contrast matrix multichannel texture classification through modification measure [3]. This policy is much robust when the textures of the kernels and the connection weights by a are unstable, or when the difference of the texture to be backpropagation-based training rule. The first layer segregated is slight. In such cases, an adaptive strategy for units working as the convolution kernels are constrained both feature extraction and classification will be in need. to be an array of Gabor filters, which achieves a most If the filtering is computed as a two dimensional efficient texture feature localization. The following layers convolution of the image and a kernel with a finite work as a classifier of the extracted texture feature support, this process (and the post filtering classification) vectors. The capability of the KM Net and its training can be incorporated into a layered neural network rule is verified with basic problems of synthetic and architecture with a local receptive field. An integration of fabric texture images, and also with a biological tissue the convolution filter into a trainable neural network has classification problem in an ultrasonic echo image. been done by Jain et al. [4]. There, the connection weights between the input layer and the first hidden layer working