Tommy W. S. Chow and M. K. M. Rahman

Image Classification is a challenging problem of computer vision. Conventional image classification methods use flat image features with fixed dimensions, which are extracted from a whole image. Such features are computationally effective but are crude representation of the image content. This paper proposes a new image classification approach through a tree-structured feature set. In this approach, the image content is organized in a two-level tree, where the root node at the top level represents the whole image and the child nodes at the bottom level represent the homogeneous regions of the image. The tree-structured representation combines both the global and the local features through the root and the child nodes. The tree-structured feature data are then processed by a two-level self-organizing map (SOM), which consists of an unsupervised SOM for processing image regions and a supervising Concurrent SOM (CSOM) classifier for the overall classification of images. The proposed method incorporates both global image features and local region-based features to improve the performance of image classification. Experimental results show that this approach performs better than conventional approaches.

After image segmentation, the image is decomposed into a number of homogeneous regions. In Fig. 1, it shows that the image is represented by a two-level tree, where the root node represents the whole image and child nodes represent the region-based objects. The root node is assigned to the global feature, which is the color histogram in this case. Local region-based features, such as color moment, texture, size and shape, are assigned to the child nodes. This enables global and local image features to be integrated through a tree structure.

Traditional neural-network-based classifiers only deal with fixed-vector type input data. To solve this problem, we use the SOM’s property of dimensionality reduction for dealing with tree-structured data. The child nodes’ (regions) features are first processed by a SOM to reduce the dimension of the final input vector. Thus, special arrangement of a two-layer SOM network is used to process the tree. The bottom layer consists of a SOM network to process the child nodes, while the top layer that consists of a CSOM classifier processes the root nodes. Fig. 2 demonstrates the processing of tree-structured data. To start the training process, the bottom SOM layer is first trained by the child node inputs from all tree data. After the completion of training, each child node is associated with its best-matched neuron on the SOM. Positions of these neurons are then used for inputs encoding of the root node together with the features of the root nodes. The root nodes, the final identity of the images, are then processed with the CSOM classifier. Thus, the bottom layer SOM is used for region encoding, while the top layer CSOM is used for image classification task.

In the experiment, we have compared results with traditional sets of flat image features extracted over the whole image. In addition, we compared our proposed approach with the purely region-based approach in which the same tree-structure of our method is used without assigning global features to root node. Different feature combinations were investigated for using as region features in our approach. Table 1 summarizes only the best performing feature sets in our experiments.

To evaluate the proposed approach, an image database consisting of 1000 images was used in this study. The images were categorized into 10 classes, each containing 100 images. Fig. 3 and Table 2 summarizes some comparative results from different approaches.

Source Code : imgClassi_Code.zip