Research On Multi-scale Object Detection Based On Deep Fusion Model

Object detection as a computer technology can automatically classify and localize the objects in an image.It has been one of the research hotspots in the field of computer vision,which has been applied in many fields,including autonomous driving,intelligent monitoring,medical diagnosis,intelligent agriculture,industrial detection and so on.Although the research has a long history of decades,the performance of object detection algorithm is still unsatisfactory in the face of complex scenes.Therefore,object detection still has great values in research theory and practical application.With the rapid development of deep convolutional neural network(CNN),the self-learning of image feature and the representation relation performs well on image recognition,such as object detection.The rapid development of CNN also means the transformation of image feature extraction from artificial designing to automatic learning.Compared with the traditional hand-designed feature descriptors,the CNN can effectively combine the low-level features into more abstract high-level features to reveal the essential representation of the image,achieving remarkable performance in different target detection task scenarios.Object detection based on deep CNN model has made great progress,but it still faces many challenges in practical application.Since the task of object detection usually faces the complex background,occlusion,deformation,various scale of the objects in an image,it is difficult to perform the feature extraction.As a result,this could lead to poor detection results and weak generalization.To address the above-mentioned limitations,we have built object detection models from four aspects,including receptive field-based region proposal network,key-point-aware of bounding box,visual attention mechanism and adaptive feature fusion,and then these designed models have been applied in various scenes.The main contributions of this paper are shown as follows:1.The receptive field-based proposal generation network(RFP-Net for short)is proposed,which effectively improves the quality of object proposals..This network adopts receptive field as reference box for bounding box regression,which ignores the complex design of anchor boxes.Moreover,a positive and negative sample selection strategy based on effective receptive field region is optimized,so that the samples of different scale objects participate in the training and learning of the whole network.The experimental results on several public datasets show that the RFP-Net with different detectors can improve the accuracy of general target detection.2.A two-stage object detector based on key-points-aware has been proposed.Based on two-satge object detection framework,a corner-based region proposal network(CRPN for short)has been proposed,which can eliminate the anchors as the reference boxes to regress the location of the object.It can completely address the problem of complex parameters setting of anchor boxes.The CRPN generates a set of region proposals by predicting key-points of bounding boxes of object.A large number of experiments verify the performance of our method,and the proposed method outperforms other object detection algorithms on large-scale datasets.3.An object detection model based on visual attention mechanism has been built.The designed model addresses the question of small object detection from three aspects.Firstly,the visual attention mechanism is introduced into the backbone network for enriching the feature expression.Secondly,the sparse sampling and a new training sample selection method are introduced into the region proposal network for relieving the imbalance of samples with small sizes.Finally,an adaptive feature selection model is designed to enhance the features of region of interest.Based on the constructed agricultural pest dataset,experiments is performed to illustrate the effectiveness of the proposed method in small object detection..4.From the perspective of feature fusion,an adaptive feature fusion network is proposed.The core idea of the model is to adaptively filter the useless information from other layers through network learning,and only retain the useful information for object recognition,so as to improve the ability of feature expression.Besides,we develop a feature enhancement module,which can be used to improve the discernibilit·y of object features in top-most level of feature pyramid network.Finally,the adaptive feature network designed in this paper is implanted into the multi-stage object detector in order to further improve the result of object detection,especially the accuracy of small-scale object detection.Simultaneously,the detection accuracy of the proposed model performs well on the agricultural pest dataset with small sizes under complex background.