Table 2. The pseudocode of the proposed algorithm.

Algorithm Multi-feature fusion based on principal component analysis and Bayesian theory
1:Intput: Image I of PCB bare board defects to be detected
2:Construct 4-layer Gaussian pyramid image $\left\{I_1,I_{2,},I_3,I_4\right\}$ of defect image I of PCB bare
board to be detected
3:Calculate the multi-scale gray level co-occurrence matrix feature vector
$\text{GLCM}=\left[{\text{GLCM}}_1,{\text{GLCM}}_2,{\text{GLCM}}_3,{\text{GLCM}}_4\right]$ of I via Eqs. (1)–(8)
4:Calculate the multi-scale directional projection feature vectors
$\mathrm{pro}=\left[X_{\mathrm{RMS}},X_r,X_{\mathrm{sk}},X_{\text{mean}},X_{\text{peak}},K_u,S_F,I_F,L,I_{\mathrm{CF}}\right]$ for I via Eqs. (9)–(20)
5:Calculate the multi-scale gradient direction information entropy feature vector
Ent = [H1, H2, H3, H4] of I for I via Eqs. (21)–(25)
6:Calculating the deep semantic feature vector D = [d1, d2, d3, …, dn], n = 512 of I based
on the improved VGG-16 neural network
7:Extract features from image I to obtain M feature vectors $\left\{X_1,X_{2,},\dots ,X_M\right\}$
8:Reduce the dimension of $\left\{X_1,X_{2,},\dots ,X_M\right\}$ via Eqs. (26)–(30), obtain N feature vectors
$\left\{X_1,X_{2,},\dots ,X_n\right\}$
9:Calculate the weight $\left\{W_1,W_{2,},\dots ,W_n\right\}$ of each feature vector of $\left\{X_1,X_{2,},\dots ,X_N\right\}$ via Eqs. (31)–(38)
10:Calculate the feature vector V after parallel weighted fusion via Eq. (39)
11:Output: the feature vector $V=\left\{C_1,C_2,\dots ,C_n\right\}$ after parallel weighted fusion