Hiroshi Sakou

Knowledge-directed inspection for complex multilayered patterns

Combining the knowledge-based processing with image processing is considered a key issue in the future of visual inspection of complex patterns such as multilayered semiconductor wafers. However, present technology restricts this combination, mainly because of the exhaustively long time usually required for each type of processing. To cope with this situation, a unique knowledge-directed image processing method is proposed, in which every image processing step is controlled in real time by parametric knowledge driven by design patterns. The resulting structure of the image processor is a pipeline, in which each piece of knowledge is embodied as a combination of a hardware processing unit and control unit. In this paper the types of knowledge and their implementation are explained, and an inspection machine for logic IC wafers based on this pipelined knowledgedirected image processing is introduced.

A position recognition algorithm for semiconductor alignment based on structural pattern matching

The authors propose a position recognition algorithm that can be used for future automated production of high-density semiconductor devices. It consists of a position detection method based on a pattern structure which is expressed as the distribution of characteristic local patterns, and an adaptive thresholding method based on an evaluation index which represents the binary state of characteristic local patterns. The validity and effectiveness of the algorithm are confirmed by experimentation with actual semiconductor chip patterns. High-speed processing of the algorithm is possible by using an image processor with simple matching hardware. >

Sign language translation based on syntactic and semantic analysis

The prototype sign language translation system described herein interprets Japanese sign language to verbal Japanese language. Since sign language has a unique morphology and syntax, the sign language translation system requires linguistic translation after recognition of the input symbols. One of the main difficulties in translation is a result of the insufficiency of syntactic markers in the input information. In Japanese sign language, input is represented as a sequence of important words. In our method, semantic dependency analysis based on case frame grammar is carried out to complement the lack of information such as postpositions. Then, rule-based rephrasing is executed to obtain natural verbal language. This scheme contributes both flexibility and ease of maintenance to the translation system. A prototype sign language translation system was implemented and tested using 123 sentences extracted from a sign language textbook. Experimental results indicate that 76 percent of the sentences can be translated correctly if they have no recognition error in input sentence. However, if a recognition error occurs, the experiment indicates that the translation will become highly ambiguous. Therefore, the capability for rejecting incorrect sentences is very important, along with the capability for translating correct input sentences. Future issues include incorporation of unused information, such as facial expression, spatial information and contexts in sign language to improve translation accuracy.

Texture discrimination using self-organized multiresolution filtering

In texture discrimination the basic image processes are considered to choose effective resolution, to determine representative features and to unify subregions having the same representative features. To realize these processes, this paper presents an improved method for discriminating higher-order features by using a self-organized multiresolutional filter. First, we make a local transformation function with the maximum discrimination ability by a limited-size neighborhood operation. Next, extending the method to the multiresolutional case, we let neighbor pixels of resolution which is optimal for discrimination be chosen automatically. Moreover, we realize unification of texture regions by connecting features represented by these pixel values and by obtaining macro-features. In this paper we also describe examples in which this method is applied to extract title characters from printed pages and to extract a specific part from a map image, and show that this method using higher-order features possess sufficient discrimination ability for real texture images.

A real‐time image processing algorithm for visual inspection of semiconductor wafer patterns

The following general requirements are made in the visual inspection of the semiconductor wafer: (1) the candidates for the defect should automatically be detected from a complex multivalued pattern; and (2) only serious defects to the reliability of the element are extracted from those candidates. Such sophisticated decisions must be made with a high reliability and a high speed. This paper proposes a method for detecting serious defects as a means to cope with those requirements. The method is composed of the following two steps.

Automatic position detection method for mounting ic parts on high-density printed circuit boards

In the automatic mounting system, which mounts electronic parts on a printed board, a problem is that the object printed circuit board is becoming of higher density and the parts have a fine lead structure. It is necessary that the positions of the part and the printed board should be detected with a high accuracy. It also became important to cope with the diversified parts with different shapes. This paper proposes a new practical method of automatic position detection to solve those problems. The proposed system is composed of the position detection process for the part and the position detection process for the printed board. In the process for position detection of parts, the positions of more than one lead are detected by the pattern matching, determining the similarity between partial images, and the position and the inclination of the part are calculated based on the result of matching. The viewing field is set so that only several leads are observed by the TV camera, and the object lens is moved to X and J directions so that the viewing field can be shifted to any desired position. In the position detection of the printed board, the positions of the two marks printed on the board are detected, and the position and the inclination of the printed board are calculated based on the mark positions. To evaluate the effectiveness of the proposed method, the position detection experiment was performed for the actual part and printed board, indicating that the position can be detected with a high accuracy.