Abstract:

Objective In the automatic detection of siagle soldier camouflage, it is necessary to detect the targets at a long distance. In this scenario, the small size of the camouflaged target and the intensification of background fusion substantially increase the difficulty of detection. Therefore, a deep learning approach to tackle this challenge is proposed based on the deep learning network architecture and module structure.

Method The original dataset was extended using data augmentation and the network architecture was designed based on the BGNet model. SCNet was used for feature extraction of images, and EAM (edge-aware module) was used for detecting target edges. EFM (edge-guidance feature module) made use of the output of EAM to guide the network to locate and identify targets, NCD (neighbor con-connection decoder) was used for fusing the features from EFM output, and the CAM (context aggregation module) was employed to aggregate multi-level features to obtain the final output.

Results The quantitative results of the proposed model and the other models showed that PFNet performed poorly in this small target detection, and SINet-V2 and C2FNet had higher recognition rates but with lower recognition accuracy, indicating poor detection accuracy although they intersect with the true values. On the other hand, the BGNet model had lower recognition rates but with higher accuracy and structural similarity. The BFNet proposed in this paper was improved based on the BGNet, and after the improvement, the recognition rate was increased. At the same time, other indices measuring detection accuracy and object similarity were also improved. The proposed BFNet was found to be able to take both recognition rate and accuracy rate into account, and identify targets more accurately and comprehensively. The quantitative evaluation of the ablation experiments was carried out, and it showed that the modified EFM improved the recognition rateIby 1.35%, indicating that more targets are able to be recognized after the improvement. The modified CAM improved the recognition rate I by 0.51%, indicating that the improved CAM further improved the recognition rate I, while S, a measure of structural similarity, and the adaptive F value Fad were also hoisted, indicating that the recall rate was also improved considering the accuracy. With the modified EFM and CAM, the detection accuracy pA was slightly decreased, but the I value is improved by 1.87%. After modifying EFM and CAM, the accuracy pA was improved by 1.74% using SCNet (self-calibrated networks) as the backbone model, proving the SCNet model compensation for the decrease in accuracy caused by the improved module structure. The results of the final improvement scheme showed that the improvement rate of pA was 0.74% and the improvement rate of I was 1.35%, while the adaptive E metric Eϕadand weighted F-measure Fwβ were improved by 0.85% and 0.71%, respectively. The qualitative comparison of the proposed model with other models is shown. The baseline model could barely recognize small targets, while the improved model performs well in small camouflage target recognition task.

Conclusion The experimental results show that the proposed model performs well in the automatic detection tasks of single soldier camouflage, which indicates that the detection model in COS (camouflage object segmentation) task is suitable for single soldier camouflage detection, and the improved model offers higher the recognition rate, especially for detecting small target. The detection algorithm can be used as an aid for combatants and also provides an effective means to evaluate camouflage designs.

Key words: camouflage detection, camouflage object recognition, deep learning, small target detection, camouflage object segmentation