一种改进的基于深度学习的小目标检测方法

An Improved Deep Learning-based Approach for Small Object Detection

提出了一种改进的基于深度学习的小目标检测方法,用于解决当前主流算法针对小目标进行检测时输入图像需为小尺寸照片且模型参数过多等缺点的问题。为了解决这些问题,首先对图像进行预处理,将一张较大尺寸的图像按一定规则拆分成多张小尺寸图像后送入网络,克服了以往算法需要小尺寸图像才能进行检测的问题。对DNANet网络结构进行改进,减少其网络层数,提高了网络推断速度。使用TverskyLoss为像素分割的损失函数对损失函数进行优化,并采用渐进式学习法训练模型,使网络从普通目标到小目标的检测过程更为稳定。实验结果表明,该方法有效提升了深度学习在小目标大尺寸图像方面的收敛速度,改进后的网络对大尺寸图像的预测准确率提升了5%,预测时间缩短了25%。综上所述,提出的基于深度学习的小目标检测方法,可以方便地应用于工程实践中,并具有较高的实际应用价值。

An improved deep learning-based method is presented for small object detection to address the limitations of current mainstream algorithms.These algorithms often require input images to be small-sized photos and suffer from excessive model parameters.To overcome these challenges,a preprocessing technique is proposed where a larger-sized image is divided into multiple smaller-sized images following specific rules.These fragmented images are then inputted into the network,eliminating the previous requirement of small-sized images for detection.The DNANet network structure is also enhanced by reducing its network layers,resulting in improved network inference speed.The loss function is optimized using TverskyLoss,a pixel segmentation loss function.Furthermore,a progressive learning approach is employed to train the model,ensuring a more stable transition from detecting ordinary to small objects.Experimental results demonstrate the effectiveness of the proposed method in enhancing the convergence speed of deep learning for small object detection in large-sized images.The improved network achieves a 5% increase in prediction accuracy for large-sized images and reduces prediction time by 25%.Consequently,the deep learning-based small object detection method presented in the study can be readily applied in engineering practice,offering significant practical value.