Research And Implementation Of Multi-scale Object Detection Based On Neural Network

In recent years,advances in deep learning technology based on neural networks have made computers perform better than humans on many visual tasks,and computer vision has gradually become a popular research field.As a basic task,object detection has naturally attracted great attention.In an image,there are objects of different scales.Large-scale objects have clear details and occupy more pixels.Small-scale objects have blurred details and occupy fewer pixels.Existing object detection methods are difficult to make appropriate processing when objects of multiple scales appear in the image at the same time.Compared to images with only single-scale objects,images with multi-scale objects are more difficult to detect accurately.This thesis studies the multi-scale object detection method based on neural network,and proposes corresponding improvement methods.The limitations of approaches of multi-scale object detection mainly lies in two facts:1.The uneven distribution of objects on different scales affects label assign;2.It is difficult to generate image features suitable for describing different objects at the same time;This thesis proposes targeted solutions to these two limitationsFor label assign,this thesis proposes the Scale-Balanced loss.In multi-scale object detection,the anchor is widely used to match the real object as a training sample for the network training.Unfortunately,such a matching strategy will result in a difference in the number of matches between different objects,which will directly reduce the accuracy of small-scale object detection.Aiming at this problem,this paper proposes a loss function Scale-Balanced Loss to alleviate the imbalance of the anchor box matching process.It will balance the difference in the amount of information generated by the anchor box matching process on the loss function level,so that the overall optimization direction of the model is towards a more balanced state.On multiple multi-scale object detection datasets commonly used in academia,the scale balanced loss can significantly improve the performance of multi-scale object detection,especially for small object recall rates.For multi-scale feature,this thesis proposes a neural architecture search algorithm.For multi-scale objects,the image features required to identify them are different.Large-scale objects need their own larger detection receptive fields,while small-scale objects not only need their own feature information,but also often need to combine surrounding semantic information to be accurately identified.Most of the current object detection feature extraction networks are direct migration of networks designed for classification tasks,and few are specifically designed for multi-scale object detection tasks.It is not enough to deal with multi-scale features through feature pyramids.This paper proposes a micro-sampling neural architecture search algorithm,and designs a search space suitable for multi-scale object detection.In this search space,a multi-scale object detection feature extraction network is searched.Under the same FLOPS limit,the proposed method exceeds the state-of-the-art method.