Lightweight Real-Time Detection and Recognition Model of Intraocular Foreign Bodies Fused With a Feature Pyramid Mechanism

Accurate detection of target location and type is crucial for treating ocular trauma caused by foreign bodies intrusion. However, the traditional method of manually marking CT image targets has slow recognition speed and poor detection accuracy, which cannot meet the real-time and accuracy requirements for detecting foreign bodies in clinical diagnosis. To address this issue, we propose a lightweight detection and recognition model based on feature extraction and fusion. Firstly, the normalization-based attention module and the sigmoid linear unit activation function are introduced into the inverted residual block of the backbone network to enhance the model’s attention to salient features and improve the detection accuracy. Then, the path aggregation feature pyramid network is utilized to fuse multiscale features, enabling the information interaction between different levels of the network and enhancing the accuracy foreign bodies classification. In particular, the incorporation of the space-to-depth convolution and convolutional mixing modules into the feature pyramid network significantly reduce the computational overhead while effectively capturing the key semantic features in both space and channel directions, thereby improving the lightweight level of the model. Finally, the location and type information of the foreign intraocular bodies are obtained by this model. The experimental results demonstrate the superior performance of the proposed model in terms of mAP@0.5, accuracy, sensitivity and specificity, achieving 97.2, 93.5, 98.0 and 88.0, respectively. Furthermore, the smaller number of parameters and faster detection time allow the proposed model run in real-time on poorly configured hardware, making it more suitable for clinical applications.