基于改进Faster R-CNN图像小目标检测

Based on improved Faster R-CNN image small target detection

为了实现光照不均匀、噪声干扰下的图像小目标识别,提出了一种改进的Faster-RCNN的小目标检测方法。使用不同膨胀率的空洞卷积来代替池化层,通过扩大感受野增强主干网络的特征提取能力,同时将低层特征图和高层特征图进行特征融合,提高小目标的检测能力;其次使用k-means聚类算法生成适合本数据集的基准锚框,提高定位精度;最后在Fast R-CNN网络中引入新的特征提取层,并在损失函数中增加光照约束条件,以此学习图像光照感知特征。在自制数据集上做了对比实验,结果表明,改进后的算法检测精度提高了5.54%,验证了算法的有效性。

In order to realize the recognition of image small target under uneven illumination and noise interference,an improved Faster R-CNN small target detection method was proposed.The dilated convolution with different expansion rates is used to replace the pooling layer,and the feature extraction ability of the trunk network is enhanced by expanding the sensing field.Meanwhile,the feature fusion module is introduced to integrate the feature map of the low level and the feature map of the high level,so as to improve the detection ability of the small target.Secondly,k-means clustering algorithm is used to generate the reference anchor frame suitable for this data set to improve the positioning accuracy.Finally,a new feature extraction layer is introduced in Fast R-CNN network,and the light constraint condition is added in the loss function,so as to learn the light perception feature of image.A comparison experiment on the self-made data set shows that the detection accuracy of the improved algorithm is improved by 5.54%,which verifies the effectiveness of the algorithm.