Fig. 1. MMW imaging system （a）prototype of Sim-Image，（b）examples of contraband

Fig. 2. Diagram of MMW imaging system

Fig. 3. Examples of MMW image processing （a）、（c）2D MMW image with ground truth of difficult case 1 and 2，（b）、（d）illustration of the sampling process

Fig. 4. Analysis results of difficult case 1 intensity distribution curve of sampling point along depth direction of （a）object and （b）background，（c）comparison of the average intensity distribution curve of the object and the background area along the depth direction，（d-f）example of MMW cross-section corresponding to interval I，II and III

Fig. 5. Analysis results of difficult case 2 intensity distribution curve of sampling point along depth direction of （a）object and （b）background

Fig. 6. Comparison of intensity distribution between object and noise （a）（e）2D MMW image with sampling area，intensity distribution curve in depth direction of （b）（f）object，and （c-d）、（g-h）noise

Fig. 7. The overall framework of the proposed method

Fig. 8. Structure of intra-section context extraction module

Fig. 9. Commom structure of LSTM^［17］

Fig. 10. Illustration of our dataset （a）3D MMW image，（b）the projection in the y-direction，（c）the projection in the x-direction，（d）statistical analysis result of the dataset

Fig. 11. Structure of the ablation model

Fig. 12. The visualization results of the CAM of feature map （a）the image to be measured with ground truth，（b）the salient region of the feature map obtained by the traditional method，（c）the salient region of the feature map obtained by the proposed method

Fig. 13. Comparison of CAM visualization of feature maps of cross-section （a-b）the side view of 3D MMW image to be measured，CAM results and the side view （c）before，and （d）after Bi-CLSTM

Fig. 14. Comparison P-R curve

Table 1. Accuracy comparison with mainstream method of MMW image object detection

Table 2. Validation of Bi-CLSTM

Table 3. Number of candidate bounding boxes predicted by different method