Automatic calculation method for surface defect area of wet blue skin based on DeeplabV3+ model
Provided is an automatic calculation method for a surface defect area of a wet blue skin based on a DeeplabV3+ model, aiming at solving the problem that the automatic calculation of the surface defect area is difficult in a tannery. The method includes image data acquisition and preprocessing of a wet blue skin, training and verification of a wet blue skin surface defect segmentation model based on a DeeplabV3+ model, and automatic calculation of the surface defect area in a segmented image. The method can accurately segment common defects such as knife holes, brands and the like on the surface of a large-area wet blue skin, with a high speed of detecting a wet blue skin within 1 s.
1 . An automatic calculation method for a surface defect area of a wet blue skin based on a DeeplabV3+ model, comprising:
Step 1, using an industrial camera to acquire more than 500 wet blue skin image samples of animal skins after main tanning, wherein the animal skins comprising cowhides, sheepskins and pigskins; performing preprocessing comprising cropping and selecting, on the wet blue skin images acquired, and carrying out data annotation on preprocessed images with an annotation software to construct a wet blue skin surface defect semantic segmentation data set;
Step 2, training a Deeplab V3+ model by using the wet blue skin surface defect semantic segmentation data set created, and obtaining a wet blue skin surface defect segmentation model based on the Deeplab V3+ model;
Step 3, using the wet blue skin surface defect segmentation model based on the Deeplab V3+ model trained to conduct defect detection on the wet blue skin image, acquiring segmentation result images of different defects in the image, and calculating a proportion of each defect in the image (k i );
Step 4: calculating actual surface defect area of a wet blue skin according to following equations:
S
t
=
∑
i
=
1
n
S
×
k
i
,
S
=
4
×
H
2
×
tan
α
2
×
tan
β
2
where S is an actual area captured by the camera, H is a distance between the camera and the wet blue skin, α is a horizontal field of view of the camera, β is a vertical field of view, S t is the actual surface defect area, and k i is the proportion coefficient of different defects in the image.
2 . The method according to claim 1 , wherein the wet blue skin image samples are captured by using one or more industrial area-array cameras or industrial single-scanline cameras with more than 2 million pixels.
3 . The method according to claim 1 , wherein in Step 1, after completion of the data annotation of the wet blue skin, the wet blue skin surface defect semantic segmentation data set is divided into a training set and a verification set according to a ratio of (1.5 to 9):1, and a number of images in the training set is greater than that in the verification set.
4 . The method according to claim 1 , wherein in Step 2, in the Deeplab V3+ model, a MobileNetV2 network is used as a backbone feature extraction network, which, in combination with technologies comprising dilated convolution and spatial pyramid pooling, improves real-time performance and segmentation accuracy, so that calculation amount and parameter amount are reduced, a number of channels is increased, and more features in the image are obtained.
5 . The method according to claim 1 , wherein in Step 3, segmented defects of wet blue skin comprise, common defects comprising knife holes and brands, and the calculating a proportion of each defect in the image is not limited to one type of defects.