Research Of Object Detection Based On Multi Depth Feature Representation And Stable Center Loss

With the rapid development and advance of technologies,object detection has been applied to various applications such as face detection,video surveillance,autonomous driving and crowd counting.A more accurate and faster object detection method is required to analyze images and videos intelligently.Object detection pinpoints the location and recognizes the categories of objects in an image.Recently,approaches based on deep convolutional neural networks have achieved a good performance under different scenes,especially the region based methods(RCNN,SPP-Net,Fast RCNN,Faster RCNN,etc.).In this paper,we mainly improve the accuracy of object detection based on the Faster RCNN method,which is one of the stateof-art object detection methods.The method consists of two steps.First,locate every possible location that may contain an object.Second,recognize the category of the object in the proposals and adjust the bounding boxes for more precise localization.Although it performs well in many benchmark datasets,it still needs to be improved to perform better in complex scenes.We proposed the multi-depth feature representation to improve the object detection method.The visualization of CNN shows that features in shallower layers capture the incomplete shape of objects and plenty of local information.In contrast,features in deeper layers capture more contextual information regarding the whole shape of objects.We proposed the multi-depth feature representation to combine features from the shallow layers and the deep layers.We have explored three ways to fuse different kinds of features using ROI pooling,transformed features,and dense features.All the experiments show better performance compared to the baseline network.We also proposed a loss function(stable center loss)in order to make the feature in object detection discriminative.The Softmax loss function exists in Faster RCNN as the sole loss function to recognize object categories.Some research in similarity learning have shown that the feature learned by Softmax loss cannot ensure that the distance between the object of the same categories is closer than that with different categories.It's easy to mix up the categories of objects which have the similar background.Following the idea of center loss in classification task,our proposed method maintains the center features for objects of each category.It pulls the feature of same categories closer to its center features and increases the distance between objects with different categories.We use a smooth L1 function to calculate the distance between the features in order to reduce the effect of isolated points and make the whole training process more stable.We evaluate our proposed method on PASCAL VOC 2007 dataset,the best result achieves mean average precision(m AP)of 73.1%,which is better than Faster RCNN.We also evaluate the effectiveness of stable center loss on MNIST dataset.When we compare the features in the image-retrieval datasets,it also shows a better performance.