Detection algorithm of container surface leakage based on improved YOLOv5s

When soap bubble method was used to detect the surface leakage of closed containers, an improved algorithm based on YOLOv5s was proposed, which is called YOLOv5s-PBSI algorithm. The problems of false detection of interference bubbles and missing detection of small soap bubbles can be solved by YOLOv5s-PBSI. Firstly, Pyramid Pooling Module (PPM) is introduced to improve the discrimination ability between target soap bubble and interference bubble; Secondly, the small target detection layer is added and the feature fusion network is improved according to the Bidirectional Aggregation Feature Pyramid Network (BAFPN) link idea to improve the small target detection ability; Finally, the model was lightened using Inverted Residual Block (IRB) to reduce the model parameters quantity. Several experiments and ablation analyses have been performed on the self-built soap bubble dataset to compare the efficacy of YOLOv5s-PBSI with different state-of-the-art target detection methods. YOLOv5s-PBSI delivers an improved runtime of 0.014 s (~ 69 fps) for detecting the target soap bubble, the model parameters quantity reduced to 10.87M, and the mAP has increased to 98%, which is also competitive when compared with the latest model.