A Defect Dttection Of Transparent Plastics Based On Context-aware Saliency Detection

With the development of science and technology and the advent of the era of big data, the images containing lots of data information spring out like bamboo shoots. How to accurately identify significant parts of the image and then how to use this data for detection has become more and more important. The emergence of saliency detection is the key to this problem. The purpose of saliency detection is to detect fixed-point or highlight in the image which can catch people’s attention at first sight. Saliency detection simulates the human visual system in which eyes can observe images, the brain analyze, and then the body response. Human can response to the complex environment efficiently. Although computer cannot surpass the human vision system, it can work persistently, so it overwhelms the human vision system in the complex industry environment.It is mainly dependent on the eyesight to detect the defect of transparent plastics package. So, there will be such problems as different standard, low efficiency, big labor intensity. The simple image segmentation and feature extracting cannot detect the defect of the transparent plastics efficiently. In order to solve this problem, the saliency detection method with high speed and high precision has been proposed. The content and the innovations are as follows.1) It In order to extract the target area,the improved point Hough transform is adopted.In this paper we adopted two kinds of superpixel segmentation, in which one is SLIC segmentation, the other is graph-based segmentation. In view of the SLIC segmentation, the BP neural network algorithm is adopted to classify the superpixel block. For the graph-based segmentation, the advanced context-aware saliency detection is adopted to detect the salient regions.2) In this paper, we compare the merits and demerits of these two methods. At last, we adopt the improved context-aware saliency detection. Due to the shape characteristics after the saliency detection, we use principal component analysis to extract 12 shape features. At last we use the SVM based the binary tree which making use of the 12 features to classify the defects. The results show that it can display the defects in real time and finally realize classification and statistics. So it is convenient for users to look up and compare.3) We compare the context-aware saliency detection method with image segmentation, edge extraction, neural network and other significant saliency detection method. The experimental results show that this extraction method can effectively detect the defects of the transparent plastic package of high accuracy and short run time.