Image segmentation is also an important branch in the field of computer vision. It is often used in image target detection and image analysis. At present, image segmentation algorithm based on deep learning has become a research hotspot. For the method of this subject, using the image segmentation method based on deep learning not only has more accurate segmentation results, but also can transfer and train the parameters trained in the target detection task, so as to realize the rapid convergence of network training. Therefore, on the basis of making the bucket tooth target detection data set, the bucket tooth image segmentation data set is also established.
2000 bucket tooth images are used as the training set, including 930 bucket tooth photos of the real bucket (obtained from the target detection training set of the real bucket), 1000 bucket tooth photos of the broken model bucket (due to the difficulty of obtaining the broken tooth image) and 70 bucket tooth photos of the normal model bucket. The verification set used 600 images, 430 target detection test sets from real buckets, and 170 broken teeth from model buckets. Here, the data used is named as the STIs (Shore tee for image segmentation) dataset.
As shown in Figure 1, labelme is used to complete manual annotation, mark multiple key points to form the object outline, and add labels at the same time, so as to generate data in JSON format. Then run the conversion command provided by labelme to get the 8-bit pseudo color map required by the training model, as shown in Figure 2.
Only the negative teeth image set is set as the segmentation data set.