基于改进Focus的小目标检测技术

为解决深度神经网络模型小目标检测效果不佳的问题，对Focus结构进行改进，提出一种即插即用的轻量级结构Focus+。搜集相关图像并建立包含5类目标的军事小目标数据集，将Focus+插入常用的目标检测模型，使用不同尺度的输入进行了多组对比实验。实验结果表明：引入Focus+模块后，模型检测的平均精度均值平均提高了0.8 %，说明其能够在占用较少算力的同时提取到浅层网络的高分辨率特征，有效提高小目标的检测精度。

Small Target Detection Technology Based on Improved Focus

In order to solve the problem of poor performance of deep neural network model in small target detection, the Focus structure is improved, and a plug-and-play lightweight structure Focus + is proposed. Relevant images are collected and a military small target dataset containing 5 types of targets is established. Focus + is inserted into the commonly used target detection model, and several groups of comparative experiments are carried out using different scale inputs. The experimental results show that the average accuracy of model detection is improved by 0. 8% after the introduction of Focus + module, which indicates that it can extract the high-resolution features of shallow network while occupying less computing power, and effectively improve the detection accuracy of small targets.