基于改进DenseNet的水果小目标检测

针对自然环境中小目标水果的检测精度普遍较低的问题，提出基于DenseNet改进的水果目标检测框架. 构建以DenseNet为核心的多尺度特征提取模块，在DenseNet不同层级的稠密块中建立特征金字塔结构，加强网络层特征复用. 结合低层特征的高分辨率和高层特征的高语义性，实现准确定位和预测小目标水果存在的目的. 引入软阈值非极大值抑制（Soft-NMS）算法，改善簇状果实结构中检测框被误剔除的情况. 与常用的Faster R-CNN网络相比，所提出的框架在苹果、芒果和杏3个数据集中的平均检测速度大于40 FPS，F1值分别为0.920、0.928、0.831，实现了检测效率及精度的提升.

Detection of small fruit target based on improved DenseNet

An improved fruit detection framework based on DenseNet was proposed aiming at the problem that small fruit target detection always obtains low accuracy in natrual environment. A multi-scale feature extraction module was built with DenseNet. A feature pyramid structure was used in dense blocks at different scales of DenseNet in order to strength the network layer feature reuse. Low-level features with high resolution and high-level features with high semantics were combined to achieve accurate localization and prediction of the existence of small fruits. Soft non-maximum suppression (Soft-NMS) algorithm was introduced to avoid the case that detection boxes were mistakenly removed in the clustered fruit structure. In three datasets of apple, mango and almond, the detection speed came up to 40 FPS, and the F1 score reached 0.920, 0.928 and 0.831 with the proposed framework. The detection efficiency and accuracy were improved compared with the commonly used Faster R-CNN network.