Author Affiliations
Information & Control Engineering Faculty, Shenyang Jianzhu University, Shenyang, Liaoning 110168, Chinashow less
Fig. 1. Weld recognition system[9]
Fig. 2. Schematic of convolution calculation
Fig. 3. Activation function images. (a) Sigmoid; (b) RELU
Fig. 4. Image acquisition device
Fig. 5. Convolution network structure used in experiments
Fig. 6. Making labels for weldment 1
Fig. 7. Computing result
Fig. 8. Recognition rate comparison among models
Fig. 9. Second weldment
Fig. 10. Label of weldment point
Fig. 11. Polygon diagram of mixed test results
Type of layers | Parameter |
---|
Convolution+activation function+pooling | C64(3×3,S=1)+RELU+Pmax(2×2,S=2) | Convolution+activation function+pooling | C128(3×3,S=1)+RELU+Pmax(2×2,S=2) | Convolution+activation function+pooling | C256(3×3,S=1)+RELU+Pavg(2×2,S=2) | Convolution+activation function+pooling | C512(3×3,S=1)+RELU+Pavg(2×2,S=2) | Convolution+activation function+pooling | C512(3×3,S=1)+RELU+Pavg(2×2,S=2) | Full connetion+activation function | 400 neutral units+RELU | Full connetion+activation function | 400 neutral units+RELU | Loss function | Euclidean distance |
|
Table 1. Network structure and hyper-parameter list
Without compression | After compression | Compressionratio | Speed ratio | Loss /% |
---|
Maximum memoryrequirement /MB | | | | Modelsize /MB | Maximum memoryrequirement /MB | Modelsize /MB | 38 | 55.3 | 1.1 | 2.33 | 23.8 | 26.3 | 1.4 |
|
Table 2. Model size and memory requirements
Method | Accuracy /% |
---|
GS | 94.3 | SURF | 92.4 | HC | 90.1 | LSE | 87.6 | GH | 93.3 | GXN | 97.5 |
|
Table 3. Test accuracy of weldment 1
Method | Accuracy /% |
---|
Secondweldment test | Mixeddata test |
---|
GS | 93.3 | 90.1 | SURF | 94.4 | 89.6 | HC | 87.5 | 86.7 | LSE | 94.3 | 88.2 | GH | 91.2 | 90.1 | GXN | 96.7 | 96.3 |
|
Table 4. Test results of multiple weldments
Method | Time /ms |
---|
GS | 35 | SURF | 27 | HC | 17 | LSE | 39 | GH | 37 | GXN | 36 |
|
Table 5. Time consumption of each method