Constructing an automatic object-recognition algorithm using labeling information for efficient recycling of WEEE

Abstract

In an attempt to select an efficient recycling process for waste electrical and electronic equipment based on the value of individual products, we are engaged in the development of an automatic object-recognition system for discarded equipment. As part of this initiative, we developed a new object-recognition algorithm that uses the information from the labels on the bottoms of digital cameras discarded in Japan, which have a relatively high value. In addition, we created a program that can continuously process multiple two-dimensional digital images of the bottoms of the discarded cameras. The algorithm developed consists of the following: 1. Identifying the manufacturer using template matching with the manufacturer's logo on the label as a template image; 2. reading the model name located close to the logo using optical character recognition (OCR) processing; and 3. extracting the model-name candidates via a similarity calculation between the result of the OCR and the model-name list. After analyzing the information on the label of the discarded cameras, we carried out an object-recognition test using the images captured inside a photography box. The results demonstrated that on average, 48% of the total number of template images was necessary to identify all the manufacturers. This value varies from manufacturer to manufacturer; however, the template image with the "highest versatility" correctly matched 42% of the models of a certain manufacturer. The model-name identification for each manufacturer was successful 92% of the time on average, which indicated the effectiveness of this algorithm and emphasized the necessity of extracting the model-name candidates from the OCR result. Finally, assuming that a continuous process will be feasible in the future, a test was carried out using the photographed images of the discarded cameras moving on a conveyor belt at a speed of 0.5 m/s. The results demonstrated that the percentage of the number of template images required to identify the manufacturer was almost identical to that for static images. Notwithstanding the limitations of the image resolution (58% lower than that of the still images), the model-name identification rate was 81%.