Object Detection And Instance Segmentation Algorithm Based On Cascade Network

The application of computer vision is ubiquitous in human life.As an important research direction of computer vision,object detection and instance segmentation have become more and more important in the fields of smart cities and unmanned driving.In traditional object detection algorithms,positive and negative samples are judged based on the size of the Intersection over Union(IOU),but generally setting a relatively low IOU threshold will get too many noise samples,making the trained detector's recognition ability low,which affects the overall Accuracy.Setting a high IOU threshold will filter out too many medium-quality samples and only retain a small number of high-quality samples.As the sample size is greatly reduced,over-fitting occurs.At the same time,the difference between the recommended regional IOU threshold and the IOU threshold used by the trainer will cause a quality mismatch problem,and it will also lead to the loss of the final training network detection accuracy.In view of the above issues,this paper improves the cascaded convolutional neural network and adds mask branches to achieve instance segmentation.The research content of this article includes the following parts:(1)Because the lower IOU will introduce noise and reduce the accuracy of the detector;the higher IOU will retain a few high-quality samples,causing overfitting.Aiming at this problem,this paper proposes a parallel cascade detection network based on a cascade network.It consists of a series of detectors connected in series and in parallel.Each detector is set with an increasing IOU threshold.Higher-quality sample distributions are used to train the next-level detector,and progressively resampling reduces overfitting.(2)This article addresses the problem of quality mismatch caused by the difference between the IOU threshold of the recommended area and the detector.An incremental threshold structure is proposed for the detector.The incremental threshold structure can gradually improve the quality of the samples,and then ensure that the recommended area and the detector's IOU threshold are similar,thereby solving the quality mismatch and improving the detection accuracy.(3)The mask structure is added on the basis of the parallel cascade network in this paper.The mask unit is added to the detector at each stage to achieve instance segmentation.This method trains the next level of detection through the improved sample quality of each level.Controller,thereby improving the accuracy of the network.