Author Affiliations
School of Automation, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, Chinashow less
Fig. 1. FPN+RPN algorithm network structure diagram
Fig. 2. Partial training samples. (a) With four defects, one gap, one scratch, and two speckles; (b) with one speckle defect; (c) with two gap defects
Fig. 3. Partial test samples. (a) With one gap defect; (b) with two defects, one gap and one speckle; (c) with one scratch defect
Fig. 4. Training loss based on ZF network. (a) Stage-1 training loss of RPN; (b) stage-1 training loss of Faster RCNN; (c) stage-2 training loss of RPN; (d) stage-2 training loss of Faster RCNN
Fig. 5. Training loss based on improved ZF network. (a) Stage-1 training loss of RPN; (b) stage-1 training loss of Faster RCNN; (c) stage-2 training loss of RPN; (d) stage-2 training loss of Faster RCNN
Fig. 6. Labeling data set with labelImg
Fig. 7. Comparison of mug defect inspection results. (a) Original Faster RCNN; (b) Faster RCNN after FPN addition
Type of layers | Number of convolution kernels | Step size |
---|
Before | After | Before | After | Before | After | Conv_1/1 | Conv_1/2 | 96 | 64 | 7×7/2 | 3×3/1 | Max pooling/1 | Max pooling/1 | 96 | 64 | 3×3/2 | 2×2/2 | Conv_2/1 | Conv_2/2 | 256 | 128 | 5×5/2 | 3×3/1 | Max pooling/1 | Max pooling/1 | 256 | 128 | 3×3/2 | 2×2/2 | Conv_3/3 | Conv_3/2 | 384/384/256 | 256 | 3×3/1 | 3×3/1 | Max pooling/1 | Max pooling/1 | 256 | 256 | 3×3/2 | 2×2/2 | | Conv_4/3 | | 384 | | 3×3/1 | | Max pooling/1 | | 384 | | 2×2/2 | | Conv_5/3 | | 512 | | 3×3/1 |
|
Table 1. ZF network structure before and after improvement
Network structure | AP /% | Average detectiontime /s |
---|
| Scratch | Speckle | Gap |
---|
Faster RCNN | 85.32 | 0 | 97.03 | 0.094 | Faster RCNN and two layers of FPN | 87.86 | 81.11 | 98.30 | 0.123 | Faster RCNN and three layers of FPN | 88.32 | 82.36 | 98.51 | 0.135 | Faster RCNN and four layers of FPN | 88.56 | 82.98 | 98.76 | 0.149 |
|
Table 2. Comparison of various network structures on classification performance