The Melanoma Object Detection Based on the Improved YOLOv5s Lightweight Bidirectional Feature Fusion Network

Melanoma is a malignant skin tumor. In order to improve its detection accuracy as well as reduce its hardware requirements for medical diagnosis, a lightweight bidirectional feature fusion network based on the improved YOLOv5s is proposed. Firstly, in order to enhance the feature extraction capability from melanoma images, we integrate the SimAM attention and Ghost module into a fused module in order to replace the original feature extraction module in YOLOv5s. Then, a bidirectional feature pyramid network structure is added to enhance the network’s feature fusion performance for melanoma feature. In final, the involution neural network operator is introduced to reduce the network’s redundancy. The Experimental results show that the improved YOLOv5s achieves a 5.8% higher mean average precision compared to the original YOLOv5s model. Its accuracy of benign melanoma detection increases 6%, and accuracy of malignant melanoma detection increases 5.5%. Also, the proposed model’s size is only 4.6M that obtains a 68.1% reduction from the original YOLOv5s model size. Therefore, the improved YOLOv5s can enhance the melanoma’s detection performance and thus is more suitable for deployment on medical equipment.