The Research On Vision-based Beer Bottle Inspector

Reusable beer bottles are widely adopted in beer production. Recycled bottles probably have some defects that may cause negative consequences for production. Hence, all recycled bottles need to be cleaned and inspected before refilling. Any defective bottles must be ejected from production line. The inspection of bottles by human inspector results in low efficiency, for the inspection process is subjective and tedious. As a replacement of human inspector, the vision-based beer bottle inspector is able to perform inspection automatically with high speed and accuracy. Compared with developed country, we achieve fairly limited accomplishment in the development of beer bottle inspector. In order to satisfy the large and increasing beer production in our country, it is necessary for us to put forward a deep and systematic research on beer bottle inspector.The concept and market requirements of beer bottle inspection are introduced in this thesis. An overview of related machine vision technologies is presented. This thesis illustrates the mechanical structure and electrical architecture of beer bottle inspector and focuses on the study of the algorithms for the inspection of beer bottle. A prototype and two experimental systems with the software for study of the inspection algorithm are developed.The main contributions of this thesis are as follows.1. Referring to the advanced achievements in beer bottle inspection, a scheme of beer bottle inspector is proposed according to the practical requirements. The in-line and rotary conveyer systems are suggested. This thesis illustrates the in-line beer bottle inspector about its mechanical structure, optical and lighting system. Three types of electric control system are discussed, which are PC-based, DSP-based and vision sensor-based respectively.2. A series of inspection algorithms are proposed for the inspection of bottle wall and bottle bottom. A method based on the histogram of edge points is applied for real-time determination of inspection area. The traditional edge detection methods are tested for inspection and are proved to be unable to satisfy the inspection accuracy. Based on the analysis of experimental results, an approach using statistical characteristics of connected component and experiential rules is designed for practical usage. Two methods using double-scale analysis are adopted to improve the result of pre-processing, which is based on correlation function and fuzzy reasoning using modified output table respectively. Experimental results show that the method based on fuzzy reasoning provides a better pre-processing result.3. This thesis makes an attempt to use neural network for the judgment of defect in the inspection of bottle wall and bottle bottom. Experimental results show that a single neural network may not satisfy the requirements and some traditional methods based on multiple neural networks also achieve very limited effects. On the basis of empirical study, this thesis presents an novel method using ensemble of multiple neural networks, which solve several crucial problems in the application of multiple neural networks, including how to collect the samples for the training of component network and how to train the component network suitable for ensemble and how to select the component network to constitute a multiple neural networks using genetic algorithm. This method is proved to be more accurate in the judgment of defect than the method using experiential rules.4. For bottle finish inspection, several algorithms are suggested according to the special requirements. A method using the center of mass of the image is tested for the determination of the center of finish. This method separates the pixels of the finish from the image according to gray value of pixels or the difference between pixels. However, experiments present no satisfying results. Based on the discussion, two useful methods for the determination of the finish center are proposed. The first one uses a varying and moving circular template to approach the center of finish step by step. The second one is a compound method combing the advantages of many algorithms and mainly based on the principle of Hough transform. For defect detection of finish, two methods that using circular template and circularly scan are firstly discussed and tested with dissatisfying results. A useful inspection method based on experiential rules and the radial projection of finish image is introduced in detail. An alternative inspection method using neural networks is provided to achieve better result, in which two level neural networks are used for low-level inspection by inspecting the same point with several different input patterns and high-level judgment respectively.5. By the aids of a prototype and two experimental systems with the software developed by ourselves, many practical problems are solved and a beer bottle image database is built. Taking advantage of these advanced tools, researchers work more efficiently in the system design and study on inspection algorithm. It creates a good prospect for further study.This thesis provides satisfying solutions to most problems of beer bottle inspector. Theoretical study and practical experiments have proved the efficiency and feasibility of the algorithm for inspection presented in this thesis. Further research is supposed to benefit from the achievements of this thesis.