Object Detector Optimized For Small Object And Its Application

Object Detection is one of most popular and most challenging research direction in computer vision.The main task of object detection is location and classification of particular instance.Object detectors based on deep learning now are under a heated discussion.These detectors are commonly divided into two-stage and one-stage.Due to simple structure,fast detection speed and the facility of end-to-end,one-stage detector has been more widely studied and applied.This article tries to take a deep insight into common problems occur in one-stage detector:weak performance on detecting small objects and extreme imbalance between positive/negative samples during training.To enhance the performance of one-stage detector on detecting small objects,two different methods are proposed in this article——shallow feature fusion and attention mechanism respectively.SF-SSD(Shallow-Feature Fusion Single Shot MultiBox Detector)and AD-SSD(Attention&DSC Single Shot MultiBox Detector)are constructed based on them.Shallow feature fusion raises precision for small object through few convolution and upsample operations.Attention mechanism equips model with capability to filtrate useful vision information,especially for detail vision feature.The test on Pascal VOC07+12 shows that SF-SSD and AD-SSD improve APS(AP for small object)by 7.9%and 6.3%respectively.Due to the better performance on small object detection,mAP(Mean Average Precision)of both reach 80.1%and 81.7%,which shows a significant improvement over SSD.To solve the two problems frequently appearing in one-stage object detector:extreme imbalance between positive/negative samples during training and anchors pre-defined deeply depending on manual settings,an anchor-free efficientnet-based object detector(AEOD)is proposed in this article.AEOD first selects out the feature points that fall in the target box,then calculates the cost matrix based on values predicted by these feature points,finally assigns the positive/negative samples to the target dynamically according to the cost matrix during the training.The number of positive/negative samples therefore are balanced to enhance the performance of the model.AEOD directly predicts location and shape of the object through the feature points in the feature maps.As a result,not only the step of pre-defining anchors can be skipped,but also the number of objects detected increases.In addition,the scalable backbone(EfficientNet)improves the generalization ability of AEOD,it can therefore receive multi-scale input.AEOD achieves the highest 91.3%mAP and 39.3%APS on Pascal VOC07+12 when input size is 1280×1280.In this article,AEOD is applied to the process of aircraft maintance.The status of flight operated in the parking apron can be achieved autonomously by detecting key objects,such as the power cord attached to the nose of aircraft,refueling pipe attached to the wing,opened cargo door and wheel chocks locked nearby the front wheel.The dataset including 2474 images are collected from the real airdrome.AEOD achieves 89.4%mAP in the test.AP for key objects,such as power cord,refueling pipe,cargo door and wheel chock are all above 90.0%.