Object Detection Based On Multi-Level Feature Fusion

Object detection methods,which may combine technologies from image processing,pattern recognition and machine learning,can locate and identify objects from videos or images and it has become a fundamental supporting technology in a lot of fields.But many traditional object detection algorithms are only suitable for the simple scenarios.The newly developed object detection methods using deep convolutional neural network are more powerful in feature extraction and abstract expression,which can obtain better results and can meet the needs in complex scenarios as well.After analyzing most recent research on object detection,we made three improvements on the Faster R-CNN algorithm and verified the effectiveness of the proposed measures using the PASCAL VOC datasets.The specific improving measures and experimental results are described in following:(1)The ways of feature fusion are optimized.We systematically study the structure and properties of feature extraction network as well as object proposal generation network of Faster R-CNN.Then according to the characteristics of network structure and data,a multi-layer feature fusion method is proposed to obtain better object proposals which can finally improve the accuracy of object detection by 2.8% compared to the original algorithm network structure.(2)The mechanism for selecting of candidate bounding boxes is optimized.We use the mixed non-maximum suppression algorithm to filter and suppress the candidate regions that generated by the region proposal generation network of Faster R-CNN.Without considering the location reliability but only the classification score,the quality of the original regions generated by Faster R-CNN are low and they have a large number of invalid regions.The experimental results have shown that the accuracy of our algorithm is 0.6% higher than Faster R-CNN.(3)The super-resolution analysis technology is introduced to improve the detection of small or blurry objects.We use the sub-pixel convolutional neural network to selectively improve the detection accuracy of those objects which are hard to detect because the size of the object in the image is too small or image region of the object is blurry.The experimental results show that our algorithm achieves an accuracy of 81.3% in object detection.