Identification and counting of the reeling cocoon number per thread for the automatic silk reeling machine

Abstract

To accurately and in real-time detect the reeling cocoon number per thread, a recognition counting algorithm based on object detection is proposed. Taking advantage of the continuity of cocoon targets in multi-frame images, the Yolact algorithm is first used to pre-recognize the cocoon reeling video frame by frame and mark the cocoon target in the image. Then the kernel density estimation and frequency filtering methods are combined to screen and optimize the preprocessed images. The kernel density estimation method calculates the probability density of cocoon location points to remove misidentified points. The frequency filtering method analyzes the fluctuation of the number of pre-identified cocoons in all the images and removes the images with low frequency and non-continuous cocoon numbers. These two techniques enhance the quality of the image that will be subjected to analysis. Finally, the number of cocoon targets in each processed image is counted, and the values with a clear concentration of trend points in the statistical distribution are selected as the final results. The finished algorithm was tested with images captured in the laboratory and at the production site, and it was found that it is possible to precisely determine the reeling cocoon number per thread in real-time.