Heading 1 The comparison between the industrial scene and the natural scene is as follows:
characteristic | natural scene | industrial scene |
---|---|---|
scale | huge change | little change |
block | covered | no occlusion |
form | huge change | little change |
category | Many categories | Less category |
illumination | unstable | Stablize |
interference | Great interference | little interference |
Natural scenes generally have strong semantic information, while defect detection generally has weak semantic information, and defect detection can generally be identified by using local regions. |
Title 2 Deficiency Summary
It is easy to start with the classification of defects. Here are three induction methods:
Induction 1:
Texture defects: instead of the original sample texture expression, the position, size, and shape are not fixed; scratches, dirt, etc.;
structural defects: related to the target structure, its position and shape are relatively fixed, and there may be no concept of quantification (errors and omissions) reverse);
other defects: such as medical images, some infrared thermal imaging, ultrasonic imaging, etc., may not be able to establish an accurate correspondence with the naked eye
.
Induction 2 (from the perspective of normal sample modeling):
Texture (generally refers to repeated structures, there may be textures with relatively large particles)
Non-texture alignment: related to the structure, but can be aligned
Non-texture cannot be aligned: has nothing to do with the structure , but it is difficult to align
the above
Induction 3 (morphology):
Addition: Dirt, foreign matter, adhesion,
Subtraction: Incomplete, scratches, damage
Replacement: Mixed colors, different colors, impurities, confusion
Deformation: Distortion, size, folds
Title 3 Feasibility Analysis
Obvious: Defects are clearly visible, easy to identify with the naked eye, and also a requirement for optical imaging;
Clear: Defect standards are clearly defined, without controversy, and are for screening requirements;
Title 4 Data Difficulties
Difficulty, diversity, imbalance, dirty data.
(1) Difficult to separate data: easy-to-separate samples (that is, obvious defects and obvious non-defects) cannot be wrong; missed detection and false detection are balanced; (2) Insufficient diversity: it is difficult to collect all
types of defect samples "Description", combining normal sample learning and data generation methods to reduce the impact of "insufficient diversity";
(3) Sample imbalance: sample level imbalance, a large number of normal samples, NG samples account for a small proportion; defects account for a small overall, It leads to time-consuming and difficult to control false detection; the category is unbalanced, a certain type of defect accounts for a large proportion, and some account for a very small proportion, which can be solved based on a large number of samples.
Title 5 Data Dirty
Dirty data means that the labeling category is mistaken when labeling. Dirty data will have an adverse effect on network training, and forced training will have the risk of overfitting. Because the network extracts general features, it cannot fit defects and can only fit other noises.
Dirty data is easier to handle. In the final analysis, it is a problem of data labeling.
The most complete industrial surface defect detection data set and thesis chicken open source project in the whole network: https://github.com/Charmve/Surface-Defect-Detection/Learn
to know about Peter Pan, organize it here! ! !
https://zhuanlan.zhihu.com/p/375828501