CAPITULO 2: POTENCIAL DEL PAÍS MERCADO
2.2. Perfil país
2.2.1. Aspectos culturales
In this chapter, a novelquery by low-quality image technique which uses a binary match- ing algorithm together with pyramidal wavelet transform is proposed. A simple but highly time consuming technique, the pixel matching technique is also developed, to be use as a yardstick for the evaluation of the proposed QBLI algorithm. The pixel match- ing technique gave very good retrieval accuracy, while the proposed QBLI algorithm gave a comparable performance. This implies that the QBLI algorithm is almost as good as the pixel matching algorithm, but has the advantage of being much faster computation- ally, making it more suitable for interactive use. The novel QBLI method illustrates the importance of the wavelet-based feature extractor in image analysis, where in PWT, we have a very fast algorithm and compact feature vectors. These two constraints are important in using the proposed QBLI method where multiple feature extractions as well as multiple feature vectors are necessary for each database image. Other feature extraction techniques are either slow or have large feature vectors.
Another important observation from the experiments conducted is that the binary image matching and thresholding method proposed is an excellent way to search using low- quality images. This is shown by the fact that the basic PWT technique gives a very poor retrieval accuracy when used for this kind of problem. This suggests that other general purpose feature extraction methods might also fail to deliver a good performance if not accompanied by a special pre-process like the binary image thresholding and matching technique. Further experiments suggested the use of Manhattan distance as the distance classifier in order to get the best results. Normalising the distance function appears to over-amplify the noise associated with the queries and therefore is not suggested. The choice of decomposition levels is also crucial as the retrieval performance decrease with lower number of decompositions. Since decomposing an image without imposing a limit might de-stabilize the features, 4×4 resolution is taken as the smallest possible sub- image size. Hence the best number of decomposition levels is found to be 6, resulting in 19 features.
The choice of wavelet basis is not very crucial although from the experiment, the best basis was found to be Coiflet 6-tap wavelets. The binary wavelet transform however is not suitable for use with the proposed algorithm as it does not carry enough information to perform good discrimination. The optimum number of binaries to be created for each image is found to be either 49 or 99. Increasing the number beyond 99 does not appear to improve the results, while reducing it below 49 reduces the performance quite drastically. If speed is not a crucial factor, it is better to use 99 binaries as it gives the best performance, otherwise using 49 binaries offers a comparable result. Finally the proposed algorithm lends itself quite well to solving query by other low-quality image types. Experiments showed that good results are obtained in solving images of inappropriate brightness and contrast, highly compressed images, low-resolution images,
quantized images and noisy images.
Although the proposed algorithm gives good performance in solvingquery by low-quality image, it still offers plenty of room for improvement. For example the system might face a problem if the object to background ratio of the image differs significantly from the target images. In this case, a special pre-processing algorithm to detect plain background can be developed so that all images will have their default object to background ratio, hence solving the above problem. The accuracy of the QBLI algorithm can also be further improved if some sort of text annotation can be associated with it, i.e. to use both text-based and content-based image retrieval (TBIR and CBIR) as in Figure 2.1. The TBIR can act as a pre-filter for the content-based retrieval process applied at later stage.
Texture Feature Extraction
In this chapter, texture representation schemes are reviewed and an evaluation of several texture features is conducted and they are compared with wavelet-based texture features. A novel implementation of the discrete wavelet frames is introduced and evaluated for texture extraction.
4.1
Introduction
Texture analysis is one of the main topics in computer vision. There is a large volume of research on different approaches and it is not possible to cover all of them in this literature review. Several popular methods and some other new approaches are chosen as a representative cross section to review. These methods belong to statistical, model- based, signal processing and structural categories. The reason why structural methods are not considered in this thesis is because they possess a number of impracticalities for our application; for example some have many pre-processing stages in order to extract texture primitives, while most of them are computationally intensive, and the assump- tion is that the input images contain strong regular features. There are several issues that this chapter is concerned with:
• Computational intensity. Although a lot of methods are being developed and have a high classification rate, the computation intensity needs to be considered. The goal of this thesis is to develop a texture matching algorithm for an interactive content-based retrieval application. Such a method should be as rapid as possible to avoid the user waiting for the query result. Furthermore, the number of images in the museum database for this thesis is in the region of tens of thousands of images, and could be up to hundreds of thousands in the future, which will requires a lot of time to extract features from all of the images. However this factor is not so crucial since it does not involve interaction with users.
• Properties of algorithm. It is important to identify whether a texture matching technique supports translation, rotation and scale invariance. Nearly all the ap- proaches are translation invariant, and some of them include rotation invariance. However, there are very few studies on rotation and scale invariant texture features, since scale is a difficult task to solve in the context of texture.
• Number of textures for testing. A number of texture papers only evaluated a small subset of Brodatz textures which are treated as a standard testing set by researchers. However, evaluating only a few textures can cause a number of prob- lems:
– It is very difficult to find the best method based on results reported in the papers.
– There are some textures in the Brodatz set that are known to have a low retrieval rate. We believe that a method should be tested on the entire set to see the results on the mixtures of different kinds of patterns, since in practice we should not expect an image database to always have clear and well defined textures.
– The developer would find it difficult to choose a texture method to suit his or her own purpose. For example, a person may want to develop a technique that has the best result on lace patterns from the Brodatz set, disregarding other texture result.
• Classification vs retrieval. Most texture analysis papers evaluated their texture features based on texture classification or segmentation. Since our application is image retrieval, this may lead to inaccurate performance measures if the texture methods are evaluated based on their classification ability. In this thesis, all tex- ture retrieval performance is evaluated in terms of precision and recall. Based on information retrieval theory, precision is defined as the fraction of images in the answer to the query that are relevant, and recall is defined as the fraction of the relevant images in the answer to the query. They are calculated as below:
P recision= Number of retrieved images that are relevant
Total number of retrieved images (4.1)
Recall= Number of relevant images that are retrieved
Total number of relevant images (4.2) Precision and recall are normally represented on a graph on apposing axis. This data is created by calculating the precision and recall up to each of the rankings in turn, so forming a graph containing as many points as documents retrieved. A perfect retrieval where only relevant documents are retrieved would form a straight line. Examples of a precision-recall graph are shown in Figure 4.1. By calculating the area under the graph, we obtain the average precision for the query. Recall is
also sometimes termed as recognition rate, and is also used individually to measure the progress of the retrieval rate against the number of retrieved documents. In this thesis, to compare the retrieval performance precisely, we will refer to the precision-recall plot and the recognition rate at several distinct places.
Pr ec isio n Recall 0 0 1 1 Pr ec isio n Recall 0 0 1 1
Figure 4.1: (left) Precision-recall graph example (right) Perfect precision-recall graph
The following section describes several popular texture feature methods, including the three different wavelet-based methods. After the review, a retrieval performance exper- iment is undertaken and evaluated between several selected texture features.