与常规目标相比,小目标在图像中的像素占比较小且特征不明显,而基于卷积神经网络的目标检测算法对小目标的鲁棒性较差。因此,基于改进的多尺寸反卷积单次多目标检测(MDSSD)网络提出了一种多特征融合小目标实时检测算法,可在不影响实时性的前提下增强算法对小目标的检测能力。为了提高算法的实时性,删减了MDSSD网络中的一部分融合模块,并用双线性插值操作代替特征图尺寸变换过程中的反卷积操作。为了提高算法的准确率,在特征图预测阶段,对每个预测层添加残差预测模块。针对算法中存在正负样本严重失衡的问题,用Focal Loss作为分类损失函数。实验结果表明,本算法在PASCAL VOC2007数据集上的平均精度均值(mAP)为81.7%,在自制轮胎瑕疵数据集上的mAP为90.6%,每秒检测帧数为50 frame。

Compared with common target, small target occupies smaller proportion of pixel in the image, the feature is not obvious, and target detection algorithm based on convolutional neural network is less robust to the small target. Therefore, this paper proposes a small target real-time detection algorithm based on multi-scale deconvolutional single shot detector (MDSSD) network, which can enhance the algorithm's ability to detect small targets without affecting the real-time performance. In order to improve the real-time performance of the algorithm, a part of the fusion module in the MDSSD network is deleted, and the deconvolution operation in the process of the feature map size transformation is replaced by a bilinear interpolation operation. In order to improve the accuracy of the algorithm, in the feature map prediction stage, a residual prediction module is added to each prediction layer. Aiming at the problem of serious imbalance between positive and negative samples in the algorithm, Focal Loss is used as the classification loss function. Experimental results show the mean average precision (mAP) of this algorithm on the PASCAL VOC2007 dataset is 81.7%, the mAP on self-made tire defect dataset is 90.6%, and 50 frames per second is detected.