Structure of the IRNN

    The image recognition neural network (IRNN) described here consists of a bottom layer, a top layer, and one or more intermediate layers depending on the complexity of the input images. The architecture of the IRNN is divided into two segments: visual and associative.

    Fig. 4.13 shows the structure of the simple IRNN. The network consists of three layers. Let us assume that a task is to recognize four different images shown in Fig. 4.14. Each image is divided by the neural network into four subimages, S1; S2; S3, and S4. So the first layer consists of four blocks (B1; B2; B3; B4), these blocks operate in a self-organized mode to classify the subimages. For instance, the input patterns for B1 are two portions of a circle and two portions of a rectangle as shown in Fig. 4.15. The resulting classes, for blocks B1 through B4, and their outputs are shown in Fig. 4.16. The outputs from blocks B1 and B2 are combined to form a feature which will be classified by block B5. Also, the outputs of B3 and B4 form a feature for block B6. Fig. 4.17 shows the classified input patterns and output codes for blocks B5 and B6. Finally, the last layer learns the relation between the combined local features and the desired recognition. The resulting relations for the last layer are shown in Fig. 4.18.