| dc.description.abstract | The use of methods based on computer vision to detect product defects is an issue that has been widely discussed for a long time. It is characterized by the efficiency of production and the degree of automation. Furthermore, it can also replace human eye detection in some environments that are not suitable for manual work. In addition to improving efficiency, it will not cause visual fatigue as human eye after a long time using. It can maintain better efficiency and quality in a long-time repetitive works.
To solve the problem of detecting tiny defects on object surface, we adopt the deep learning methods. Deep learning is one kind of algorithm that imitates the structure of neural networks. Before training, parameters in the model are all randomly initialized values. It requires a large amount of training data and a fine-defined loss function to make network learning. Step by step, adjusts the parameters in every iterative training, and finally obtains a model that can give the expected outputs according to the input.
The goal of this research is to find the defective region on object surface using semantic segmentation. We propose our symmetric model based on U-Net. The main ideas to make encoder and decoder stronger, and add self-attention mechanism to the model to further improve the learning effect. We still keep the model in a symmetric structure, but use the residual blocks as a replacement of continuous convolutional computing without skip connection. More, we adjust the number of convolutional layers to fit the task of detecting tiny defects on object surface. Self-attention mechanism is applied in two part of our model where high-level feature enhancement and the merging of high-level and low-level feature while feature up-sampling. In particular, the procedure of high-level feature enhancement, has been added in the encoder output part. Two self-attention modules have been used to perform position and channel self-attention enhancement respectively. The other part has been added in the up-sampling step of the decoder, the module performing self-attention to enhance the low-level features before concatenating with the high-level feature.
Compared with U-Net, our model not only does not increase much hardware cost, but also improves the ability to detect defects of small regions on object surface. In experiments, we choose the keycap images as target of our surface defect detection. 594 images are picked for training and 65 images for testing. We compared many different network architectures and self-attention modules. The residual symmetric encoder and decoder enhance 1% of recall. The position and channel attention module raise up recall for 4% and 2% respectively. The global attention upsample increase 2% of recall. Finally, the final version of our defect segmentation network achieve 85% MIoU and recall. | en_US |