Research On Object Detection Algorithm Of Multi-Receptive Field Branch Network Based On Cascade Structure Improvement

Object detection is an important research direction and hotspot in the field of computer vision,which has been widely used in unmanned vehicles,robots,video surveillance,pedestrian detection,vessel detection and other fields with the task of classify and coordinating the objects from the complex scenes.The current mainstream algorithm is the object detection algorithm based on deep learning,but there exists the problem of poor detection effect on small objects.Aiming at this problem,most of the existing improved algorithms use the method of feature fusion to enhance the detection effect of the algorithm on small targets.However,this method has added many network parameters,resulting in a slow detection speed.Therefore,this paper proposes an object detection algorithm of multi-receptive field branch network based on cascade structure improvement,which improves the overall detection effect,especially the detection effect of small targets,and has a high detection speed.Here are main work and contributions about the object detection algorithm.1)This paper proposes a multisensory field branch network.From the perspective of the receptive field,the cavity convolution is introduced,which improves the accuracy of the overall objects detection.In the training phase,the improved classification Loss function—Enhanced Focal Loss was proposed to replace the original cross entropy Loss function.Therefore,the network training is more obviously biased towards the difficult-to-classify samples,which improves the training effects.2)This paper designs two enhancement detection modules to cascade with the original network.For one thing,STFCN(Small Target Feature Combination Network)reinforcement module of small objects detection is added into the small object detection.The effect of the algorithm on small target detection is improved.The large target detection design is integrated into the large target detection module of AF,which improves the detection effect of the large target.The improved algorithm in this paper is compared with several mainstream object detection algorithms on the PASCAL VOC 2007 data set,which conveys the effectiveness of the proposed algorithm.