The quality of cocoons directly affects the quality of raw silk. Before reeling, the cocoons should be sorted to remove the inferior cocoons that cannot be used for reeling. The double cocoon is the best raw material for making silk quilts. Therefore, it should be separated from other types of inferior cocoons and classified separately in the process of selecting cocoons. The traditional cocoon sorting basically relies on manpower, which involves high labor intensity, low efficiency of cocoon selection and high missing rate, and the accuracy of cocoon selection is easily affected by subjective factors such as the technical level and mental state of cocoon selection personnel. With the rapid development of artificial intelligence, machine vision technology is widely used to replace human eyes for detection and recognition. At present, in machine vision technology, the recognition of double cocoons mainly includes area threshold method and the threshold method of the ratio of long to short axis of ellipse. The disadvantage of the area threshold method is that small double cocoons cannot be detected. The threshold method of the ratio of long to short axis of ellipse is not applicable to the double cocoon of the head-to-tail type and relies too much on the accurate fitting of the cocoon shape. However, due to the slight irregularity of the cocoon shape, the attachment silk and the adhesion between the cocoons, the ellipse fitting is different from the actual shape, which may result in people’s misjudging the cocoon type near the threshold. Therefore, there are some limitations in the identification of double cocoons by only depending on the area threshold or the threshold of the ratio of long to short axis of the fitted ellipse. Due to the dense distribution of cocoons in the cocoon image and the interference of attached silk, the gray scale of adjacent cocoons is similar, so it is difficult to conduct segmentation completely by traditional image segmentation methods. Therefore, the segmentation and counting problem of target cocoons should be solved first for the classification and recognition of cocoons. Aiming at the problems that it is difficult to accurately segment the adhered cocoons and the limitations of the existing double cocoon recognition methods, we proposed a series of algorithms for the segmentation and counting of cocoons, the shape restoration of cocoons and the recognition of double cocoons. Firstly, we proposed an improved image segmentation algorithm based on variable segmentation threshold by combining with morphological operation. Otsu binarization threshold was used as the initial segmentation threshold, and binarization, morphological erosion operation and morphological expansion operation were performed on the image. The pixel area threshold and length-width threshold of single cocoons were used to distinguish the connected region. If a cocoon is a single one, it is marked and counted. If it is determined to be connected to other cocoons, the segmentation threshold is increased again for another segmentation. Then, a cocoon shape restoration method based on edge detection and ellipse fitting was proposed. The pixels of each cocoon that had been segmented were traversed. The Canny operator was used to detect the edge of the original image. The pixel points of the segmented cocoon were used to determine the edge points of the cocoon, and the edge points were elliptically fitted to obtain the image of the cocoon. Finally, the ratio of long to short axis of the cocoon and the area parameters were calculated, and the double-parameter threshold was used to recognize the double cocoon. And then, the serial algorithm proposed in this paper was used to segment cocoon graphics, restore cocoon shape, and distinguish double cocoons in 10 cocoon images. The results show that the 222 cocoons of all the ten pictures are completely segmented, and the counting accuracy is 100%. A total of 46 double cocoons in the pictures are detected, and the classification accuracy of double cocoons and single cocoons is 98.6%. This method detects the small double cocoons which are easy to be missed by the area threshold method, and accurately identifies the cocoon type which is close to the threshold of the ratio of long to short axis. The series algorithm of cocoon target segmentation and double cocoon recognition proposed in this paper can effectively process the machine vision image to get the conclusion whether each cocoon is a double one. In the series algorithm, the pre-experiment is used to determine the area of the normal cocoon and the length of the long axis of the normal cocoon as the prior conditions in the captured image. The cocoon target with slight overlap can be separated, but the cocoons with serious overlap may be mistakenly divided into the same group, thus affecting the classification and recognition of the cocoon. Therefore, this method is suitable for the image processing of cocoons laid flat in the field of vision. [ABSTRACT FROM AUTHOR]