Research And Implementation On Image Detection Of Small Objects

Object detection is one of the basic capabilities in the field of computer vision.It is the starting point for intelligently understanding and recognizing the world,and it is also the basis for in-depth research on some tasks such as object tracking and instance segmentation.In recent years,with the rapid development of deep learning and convolutional neural networks,object detection has been widely adopted in many industries and fields.However,the effect of small object detection is not satisfactory.Whether the experimental results on public datasets or the application in life,the detection accuracy of small objects is relatively poor,far less than that of large and medium objects.However,its practical significance is very important.Many strict detection scenarios need to rely on the ability of small object detection to achieve their goals,which has also become a popular research topic.In view of the problems existing in small object detection,researchers have proposed a series of object detection methods based on convolutional neural networks and deep learning,which have a certain improvement in accuracy and speed.In view of the problems existing in small object detection,researchers have proposed a series of target detection methods based on convolutional neural networks and deep learning,which have a certain improvement in accuracy and speed.However,there are still the problems of inaccurate attention to key regions,and the inability to capture more adequate features in relation to the global context information.These problems have a certain impact on the feature extraction of small object.to help the information extraction of small objects.Based on the above two problems,this paper proposes a two-layer attention mechanism module called DoubleS-AM,and proposes related network models.In addition,this paper designs and implements a prototype system suitable for three kinds of object detection.The main optimization and improvement works include the following three items.Based on the above two problems,this paper designs and proposes a dual attention module DoubleS-AM and related network models,in addition,we designs and implements a prototype system suitable for three object detection application scenarios.The main optimization and improvement works include the following three itemsFirst,the double-attention module,DoubleS-AM,designed and proposed in this paper includes two sub-modules:the scale information extractor called SPP-AM and the deep semantic fuser called SW-AM.The scale information extractor SPP-AM is optimized by using the channel domain attention mechanism to adaptively focus on multi-scale objects,which is helpful for object detection with large differences in size.The deep semantic fuser SW-AM uses the spatial domain attention mechanism to capture the deep contextual information,and effectively fuses the deep semantic features with the shallow features.SW-AM can help to enhance the features of small objects.Second,Combining the dual attention module DoubleS-AM and different feature pyramid fusion methods,this paper designs three end-toend network model structures.We do extensive experiments on the general datasets PASACAL VOC2007 and MS COCO2017.Compared with the baseline methods,the proposed network models in this paper have increased by 4%mAP on PASACAL VOC2007 dataset.On MS COCO dataset,which is difficult to detect and has a large number of small objects,our models have increased by about 3%mAP,and the AP of small objects have increased by 3%-5%.While maintaining fast detection,the algorithm has a significant impact on multi-scale object detection,especially for small objects.Third,on the basis of the above research results,combined with application scenarios in real life,this paper designs and implements an object detection system which is suitable for three scenes,including multi-scale general object detection,helmet-wearing detection and flying-bird detection.In this prototype system,the accuracy rate of test data in three scenarios is above 88%,which increased by more than 3%compared to the traditional algorithm.The system realizes the practical application of DoubleS-AM proposed in this paper.It has both general goals and customized applications.The interaction of the system is simple and intuitive,which is convenient for users.