Tire Speckle Interference Bubble Defect Detection Based on Improved Faster RCNN-FPN

Abstract

With the development of neural networks, object detection based on deep learning is developing rapidly, and its applications are gradually increasing. In the tire industry, detecting speckle interference bubble defects of tire crown has difficulties such as low image contrast, small object scale, and large internal differences of defects, which affect the detection precision. To solve these problems, we propose a new feature pyramid network based on Faster RCNN-FPN. It can fuse features across levels and directions to improve small object detection and localization, and increase object detection precision. The method has proven its effectiveness through cross-validation experiments. On a tire crown bubble defect dataset, the mAP [0.5:0.95] increased by 2.08% and the AP0.5 increased by 2.4% over the original network. The results show that the improved network significantly improves detecting tire crown bubble defects.

Keywords: computer vision; deep learning; feature pyramid network; object detection; tire defect detection.

Figure 1

Tire Speckle Interference Bubbles. (a–f) are the bubble part of the tire laser speckle interference bubble defect image, the pixel size of these images is 67 × 67, and the size of the bubble defect is lower than 32 × 32, which belongs to the size range of small objects.

Figure 2

Improved bubble defect detection algorithm. It includes four parts: Backbone, Tyre-Feature Pyramid Network, Region proposal, and Region prediction.

Figure 3

Feature Pyramid Network—including a bottom-up feature extraction, a top-down feature fusion, and lateral connections.

Figure 4

Tyre—Feature Pyramid Network—including a bottom-up feature extraction, a multi-directional, cross-level feature fusion, and lateral connections.

Figure 5

Bounding box coordinates and size distribution plot. Statistical image of the number of bubble defects (upper left); All bounding box size images (top right); Bounding box relative position distribution image, x and y are the coordinates of the relative position of the bubble in the image. (bottom left); the height and width of the bubble (bottom right).

Figure 6

Speckle interference tire bubble defects example. These are some instances in the dataset, and the places with higher gray values in the image are bubbles.

Figure 7

Learning rate settings. The learning rate increases linearly to 0.02 for the first 500 iterations and decreases to 10% at epoch 16 and epoch 22.

Figure 8

Schematic diagram of the cross-validation experiment. The ratio of training set to test set is 7:3, and k-fold cross-validation is performed, k = 5.

Figure 9

Tire bubble defect images and detection results. The upper part is the picture with defects, and the lower part is the corresponding detection result, including the predicted defect location, boundary, and probability.

Figure 10

Detection results of FPN.

Figure 11

Detection results of TY-FPN.