Semantic event and slow motion replay extraction for sports videos have become hot research topics. Most researches analyze every video frame; however, semantic events only appear in frames with scoreboard, whereas replays only appear in frames without scoreboard. Extracting events and replays from unrelated frames causes defects and leads to degradation of performance. In this paper, a novel framework is proposed to tackle challenges of basketball video analysis. In the framework, a scoreboard detector is first provided to divide video frames to two classes, with/without scoreboard. Then, a semantic event extractor is presented to extract semantic events from frames with scoreboard and a slow motion replay extractor is proposed to extract replays from frames without scoreboard. Experimental results show that the proposed framework is practicable for basketball videos. It is expected that the proposed framework can be extended to other sports.

The eye line is defined to be a horizontal line passing the two eyes of a human face. It can be used to help locate the true positions of eyes and face. In this paper, we propose a method to extract the horizontal eye line of a face under various environments. Based on the facts that the eye color is very different from skin color and the gray level variance of an eye is high, some eye-like regions are first located. Then the horizontal eye line is extracted based on the located eye-like regions and some geometric properties in a face. Experimental results are given to show that the proposed method is robust under a wide range of lighting conditions, different poses and races. The detection rate for HHI face database is 95.63%. For Champion face database, the detection rate is 94.27%.

Sprite coding, which can increase the coding efficiency of backgrounds greatly, is a novel technology adopted in MPEG-4 object-based coding. Due to the geometric transformation applied to each nonreference frame in the procedure of sprite coding, the generated sprite is distorted and the available view angles relative to the reference frame are restricted. These problems can be resolved by using multiple sprites. An optimal multisprite generator has been proposed by Farin et al., but the optimal method requires high computation in the sprite coding cost. To treat this disadvantage, a fast multisprite partition algorithm is proposed in this paper based on frame translations and scalings. The proposed algorithm divides a video sequence into several subsequences, and a sprite is generated from each subsequence. The experimental results show that the executing speed of the proposed method is increased by 10 to 190 times than the optimal method, with the total size of generated sprites is slightly higher and the qualities of generated sprites are preserved.

The sprite generator introduced in MPEG-4 blends frames by averaging, which will make places, that are always occupied by moving objects, look blurred. Thus, providing segmented masks for moving objects is suggested. Several researchers have employed automatic segmentation methods to produce moving object masks. Based on these masks, they used a reliability-based blending strategy to generate sprites. Since perfect segmentation is impossible, some ghost-like shadows will appear in the generated sprite. To treat this problem, in this paper, an intelligent blending strategy without needing segmentation masks is proposed. It is based on the fact that for each point in the generated sprite, the corresponding pixels in most frames belong to background and only few belong to moving objects. A counting schema is provided to make only background points participate in average blending. The experimental result shows that the visual quality of the generated sprite using the proposed blending strategy is close to that using manually segmented masks and is better than that generated by Lu-Gao-Wu method. No ghostlike shadows are produced. Furthermore, a uniform feature point extraction method is proposed to increase the precision of global motion estimation, the effectiveness of this part is presented by showing the comparison results with other existing method.

The existing audio retrieval systems fall into one of two categories: single-domain systems that can accept data of only a single type (e.g. speech) or multiple-domain systems that offer content-based retrieval for multiple types of audio data. Since a single-domain system has limited applications, a multiple-domain system will be more useful. However, different types of audio data will have different properties, this will make a multiple-domain system harder to be developed. If we can classify audio information in advance, the above problems can be solved. In this paper, we will propose a real-time classification method to classify audio signals into several basic audio types such as pure speech, music, song, speech with music background, and speech with environmental noise background. In order to make the proposed method robust for a variety of audio sources, we use Bayesian decision function for multivariable Gaussian distribution instead of manually adjusting a threshold for each discriminator. The proposed approach can be applied to content-based audio/video retrieval. In the experiment, the efficiency and effectiveness of this method are shown by an accuracy rate of more than 96% for general audio data classification.

Rapid increase in the amount of audio data demands an efficient method to automatically segment or classify audio stream based on its content. In this paper, based on the Gabor wavelet features, an audio classification and segmentation method is proposed. This method will first divide an audio stream into clips, each of which contains one-second audio information. Then, each clip is classified as one of two classes or five classes. Two classes contain speech and music; pure speech, pure music, song, speech with music background, and speech with environmental noise background are for five classes. Finally, a merge technique is provided to do segmentation. In order to make the proposed method robust for a variety of audio sources, we use Fisher Linear Discriminator to obtain features with the highest discriminative ability. Experimental results show that the proposed method can achieve over 98% accuracy rate for speech and music discrimination, and more than 95% for a five-way discrimination. By checking the class types of adjacent clips, we can also identify more than 95% audio scene breaks in audio sequence.

Rapid increase in the amount of audio data and especially music collections demand an efficient method to automatically retrieve audio objects based on its content. In this paper, based on the Gabor wavelet features, we will propose a method for content-based retrieval of perceptually similar music pieces in audio documents. It allows the user to select a reference passage within an audio file and retrieve perceptually similar passages such as repeating phrases within a music piece, similar music clips in a database or one song sung by different persons or in different languages. The proposed method will first divide an audio stream into clips, each of which contains one-second audio information. Then, the frame-based features of each clip are extracted based on the Gabor wavelet filters. Finally, a similarity measuring technique is provided to perform pattern matching on the resulting sequences of feature vectors. Experimental results show that the proposed method can achieve over 96% accuracy rate for audio retrieval.

In this paper, we will present a method to extract slow-motion replays in a soccer game video. According to our observation, a slow-motion replay always happens as a single shot and appears between two gradual transitions. Based on this fact, a video will first be segmented into individual shots and only those shots appearing between two gradual transitions are considered as candidate shots. There are two kinds of slow-motion replays. One comes from a standard camera and consists of some repeating or inserted frames. The other is from a high-speed camera with larger variation between two consecutive frames. Based on these features, an extractor is then provided to check if a candidate shot is a slow motion replay. Some experimental results are given to show the effectiveness of the proposed method.

The typical model of steganography has led the prisoners' problem, in which two persons attempt to communicate covertly without alerting the warden, that is, only the receiver knows the existence of the message sent by the sender. One available way to achieve this task is to embed the message in an innocuous-looking medium. In this paper, we propose a variation of the Quantization Index Modulation (QIM) for solving the prisoners' problem. We also propose a theorem to show that the error of mean intensity value of an image block caused by JPEG compression is bounded. The proposed method embeds the messages to be conveyed by modifying the mean intensity value, and the resulting stego-image can be stored in the JPEG format with a low quality setting. Besides, a specific pattern caused by using the QIM embedding method is also identified, and this pattern will be removed using the proposed embedding method. Experimental results and the proposed theorem show that the hidden message is error-free against the JPEG distortion under the quality setting as low as 25. Furthermore, the existence of the hidden message is not only visually imperceptible but also statistically undetectable.

In this paper, a new method for extracting primitives of a regular texture via wavelet transform is provided. Each primitive is restricted to be a parallelogram. The main work of the proposed method is to extract the two displacement vectors that form the primitive of a regular texture. A contrast-based criterion is used to select appropriate subimages obtained from wavelet transform for detecting displacement vectors. An edge thresholding method is then performed on the subimages selected to locate edges. Based on these edges, Hough transform is applied to extract the displacement vectors. The proposed method is quite efficient and can get more accurate results for displacement vectors when comparing to traditional co-occurrence matrix based methods. Synthesized textures are provided to show the effectiveness of the proposed method.

The typical model of steganography has led the prisoners' problem, in which two persons attempt to communicate covertly without alerting the warden. The general way to achieve this task is to embed the message in an innocuous-looking medium. In this paper, an object-based geometric embedding technique is proposed for solving the prisoners' problem. The main idea is to embed secret data through distorting a given object and the distorted object still looks natural. In the embedding process, the secret message is first converted into coefficients of an affine transformation. Then, the coordinates of each pixel of a selected object in the cover-image are recomputed by this affine transformation. Since these coefficients are restricted in a specific range, the transformed object looks natural. In the extracting process, a coarse-to-fine iterative search is proposed to accelerate the object location and the message extraction. Experimental results show that all transformed objects can be located precisely and the whole hidden message can be extracted correctly even if the stego-image is stored in various compression formats and rates. Furthermore, the embedded message is robust enough when the stego-image format is converted from GIF to JPEG, and vice versa.

In this paper, a color image retrieval method based on the primitives of images will be proposed. First, the context of each pixel in an image will be defined. Then, the contexts in the image are clustered into several classes based on the algorithm of fast noniterative clustering. The mean of the context in the same class is considered as a primitive of the image. The primitives are used as feature vectors. Since the numbers of primitives between images are different, a specially designed similarity measure is then proposed to do color image retrieval. To better adapt to the preferences of users, a relevance feedback algorithm is provided to automatically determine the weight of each primitive according to the user's response. To demonstrate the effectiveness of the proposed system, several test databases from Corel are used to compare the performances of the proposed system with other methods. The experimental results show that the proposed system is superior to others.

Traditional approaches for texture segmentation via wavelet transform usually adopt textural features to achieve segmentation purposes. However, for a natural image, the characteristics of the pixels in a texture region are not similar everywhere from a global viewpoint, and over-segmentation often occurs. To deal with this issue, an unsupervised texture segmentation method based on determining the interior of texture regions is proposed. The key idea of the proposed method is that if the pixels of the input image can be classified into interior pixels (pixels within a texture region) and boundary ones, then the segmentation can be achieved by applying region growing on the interior pixels and reclassifying boundary pixels. Based on the fact that each pixel P within a texture region will have similar characteristics with its neighbors, after applying wavelet transform, pixel P will have similar response with its neighbors in each transformed subimage. Thus, by applying a multilevel thresholding technique to segment each subimage into several regions, pixel P and its neighbors will be assigned to the same region in most subimages. Based on these segmented results, an interior pixels finding algorithm is then provided to find all interior pixels of textural regions. The algorithm considers a pixel which is in the same region as its neighbors in most subimages as an interior pixel. The effectiveness of this method is proved by successfully segmenting natural texture images and comparing with other methods.

In this correspondence, a fast approach to motion estimation is presented. The algorithm uses the block sum pyramid to eliminate unnecessary search positions. It first constructs the sum pyramid structure of a block. Successive elimination is then performed hierarchically from the top level to the bottom level of the pyramid. Many search positions can be skipped from being considered as the best motion vector and, thus, the search complexity can be reduced. The algorithm can achieve the same estimation accuracy as the full search block matching algorithm with much less computation time.

Many approaches for recognizing all of the 5401 commonly used Chinese characters have been proposed, but they are complex. For some applications such as cheques, receipts, addresses, in Chinese, etc., only a limited set of printed characters is needed, but they are usually require a real time response and a high recognition rate. Based on this reason, here we provide a simpler approach to meet these two requirements. The approach takes horizontal strokes, vertical ones and crossings among these strokes as the character features. A fast modified Hough transform is provided to extract these features, it satisfies an important stroke property that a dotted line must not be a real stroke. In the learning stage, an error measure is defined and an optimal stroke matching algorithm is proposed to establish an accurate random model for each type of characters. In the recognition stage, another error measure is defined and a nearly-optimal matching algorithm is presented to speed up the recognition process. It is worth mentioning that the approach does not use any thinning process. Some experimental results are also provided to show the effectiveness of the proposed recognizer.

A high-speed method for elliptical object location is proposed. Through the use of the geometric symmetry, it first locates the centers of elliptical objects and classifies the boundary points. Then, each point possible on an elliptical object boundary is used to obtain the remaining three parameters of the elliptical object. Since each boundary point is at most transformed to one point on the parameter space, the proposed method is very fast. Some experimental results are also given to show that the proposed method is better than some existing methods under the consideration of space and storage.

One of the most serious problems for vector quantization, especially for high dimensional vectors, is the high computational complexity of searching for the closest codeword in the codebook design and encoding phases. Although quantizing high dimensional vectors rather than low dimensional vectors results in better performance, the computation time needed for vector quantization grows exponentially with the vector dimension. This makes high dimensional vectors unsuitable for vector quantization. To overcome this problem, a fast search algorithm, under the assumption that the distortion is measured by the squared Euclidean distance, is proposed. Using the mean pyramids of codewords, the algorithm ran reject many codewords that are impossible matches and hence save a great deal of computation time. The algorithm is efficient for high dimensional codeword searches. Experimental results confirm the effectiveness of the proposed method.>

Proposes an algorithm, which satisfies real time process requirements and has a high recognition rate, for recognizing a limited set of multifont Chinese characters used in a special application. Stroke is one of the major features of Chinese characters. Since the aim of the proposed method is to recognize a special set of printed Chinese characters, only the horizontal/vertical strokes and the crossings among these two types of strokes are used as the features of characters. In the learning stage, a random model for each type of characters is established. In the recognition stage, the character model which has the minimal error with the input pattern is considered as the recognition result. The proposed recognizer is insensitive to noise. Some experimental results show that the method exactly provides high recognition rate and has high computing speed.>

In this paper, an edge point, line point, or curve point is called a feature point. A new approach to extract junction points and to describe feature points is proposed here. It accepts as input data a binary image resulting from a feature detector without thinning. In the binary image, each black point is first classified based on the number of lines passing through it and on a local property that the classes of its neighboring points are almost the same. Next, an aggregation method is presented to group those classified points into several segments. The orientation of each segment is kept either clockwise or counterclockwise. Conic curves are then used to describe these segments. Finally, junction points including corner points, cross points, branch points, and inflection points are located. It is worth mentioning that the proposed method does not use any thinning process and curvature information. The effectiveness of the approach is also verified by one illustrative example and two experimental results.

A novel method for curve detection based on the moment-preserving principle is proposed. The method can be used to estimate curve locations and widths to subpixel accuracy. For each 4.5-unit circle in an input image that includes a curve segment, the approach derives a parabolic equation as well as a width value to describe the curve segment. Experimental results are included to show the effectiveness of the proposed detector.>