Deep Position-sensitive Network For Object Detection

Object detection is an important research topic in computer vision and an important part of various types of increasingly intelligent systems.In the past The mechanical object detection algorithm was only suitable for the object detection problems under certain conditions,the detection accuracy is easily interfered by the external environment,and the algorithm was poor in portability.The core of the intelligent object detection algorithm is based on the image recognition algorithm,so it faces the problem of insufficient extraction of the image object position sensitivity features.At the same time,intelligent object detection training model is vulnerable to imbalance data.To improve the ability of extraction object position sensitivity features,We proposes a deep position-sensing network structure that combines different degree of position-sensitive information networks for intelligent target detection algorithms.We also designs a training method based on ensemble learning ideas with the characteristics of network structure to reduce the impact of imbalanced data.Our network extracts feature maps using a modified ResNet,and then using two fully convolutional layers to produce the structure of various sizes score maps,and proposal regions are extracted using the region proposal network(RPN),Selective Search and Edge Box.Position-sensitive information networks with different degrees calculate the position-sensitive features of the target,and use the position-sensitive network to calculate the probability that each candidate region is the target region,and finally output the position and category of the target.Based on the ensemble learning idea,through the sharing of network weights,the thresholds for finding hard negative samples are continuously updated to train the training set,and the cycle training network is used to reduce the impact of unbalanced data.At the same time,the recognition of hard negative samples is improved to complete the training of deep position-sensitive networks.Our experimental results show that the proposed method can effectively enhance the capacity to learn the object's position-sensitive information.For the same experimental conditions,we train our network with various sizes assembly on PASCAL 2007+2012 dataset and test on the PSACAL 2007 testset.Most of results is better than the single size that is included in the assembly,but since the granularity of 5 X 5 and 7×7 size is too close,the performance is similar with single 7×7 szie.The best result is 74.69%mAP with 3×3 and 7×7 size.