In-Process Intelligent Inspection of the Specimen Using Machine Vision

Abstract

Quality control is a crucial component of every manufacturing process. Quality production can be done by removing the defective pieces before reaching the packaging line. Several innovative systems have proven the use of visual input and advanced computer processing to fulfill various production goals during the last several years. Product inspection technologies based on machine vision are extensively researched to increase the quality of the product and save expenses. Computer vision and Deep learning have recently evolved, resulting in powerful data analysis tools with excellent scanning quality and resilience. Authors have attempted in this direction using such a method to detect flaws present in the dimensions of the bottles, which are traveling continually on the conveyor belt. Using pictures collected from the camera, the Yolov5 object detection method is used to localize the bottle in the image. Then, the image is passed for pre-processing, such as image cropping, image gray scaling, and smoothening of the image. The next step of this algorithm uses canny edge detection to detect edges present in the image. The image with detected edges is in the form of a binary image. All the pixels are extracted from this binary image in the form of an array. After performing some mathematical calculations on the output array, the dimensions of the bottle can be determined. The bottles were inspected for any faults in the dimensions in the manufacturing. Any bottles with flaws in the dimensions are discarded and separated from the manufactured bottles. The first step of the algorithm is object detection; here, the model has achieved the mean average precision of nearly 99.5 percent for the confidence threshold set to 50 percent to 95 percent. The following entire algorithm runs in less than 847 milliseconds. Such a high-speed algorithm allows manufacturers to increase and decrease the manufacturing speed according to their needs. This algorithm can check any shape of bottle, and this algorithm is not limited to bottles, but it can also work for any shape of object. As this model is only trained on the images of the bottles, the model cannot instantly work on the other objects, but one can use transfer learning to use this model on different object. This algorithm can also detect defects in multiple objects in the production line containing the manufacturing of multiple objects in the same line. The model can classify the objects from the production line and can also be used to classify them wherever required.