Insignificant Object Detection Based On Depth Metric TripleNet-PSO Method

Object Detection is one of the main research hotspots in the field of computer vision,attracting extensive attention of researchers in related fields,and has achieved a large number of fruitful research and application achievements.However,there are many problems in the practical application.For example,the target to be detected cannot be completely distinguished from the background area;the target is weak in significance;the edge is fuzzy;the large difference in size,shape and color.Research on detection of such non-salient objects remains a challenge.Existing object detection algorithms,depending on the candidate area,are usually sensitive to the difference between the target area and the background area.Moreover,it is difficult to meet the detection requirements when there is crossover with the background area and the edge of the detection target is blurred.Aiming at the problem of insignificant target location with fuzzy edges,this paper proposes a random search method based on depth metric.Firstly,generate randomly candidate areas in the image,and formulate relevant coding mechanisms to regard each candidate area as an Agent.Then,construct and train a deep measurement network to measure the similarity between the target example and the Agent.Finally,drive Agent to search for target by adopting PSO method.This method does not rely on the selection of candidate areas.In the detection of nonsalient objects,the detection recall rate is significantly improved compared with the existing detection methods,and the detection results are reliable and effective.Since the population has only one optimal individual in the traditional PSO algorithm,and the individual to be detected is often not unique in practical applications,an adaptive target detection method based on depth measurement is further proposed,which is used to improve the traditional PSO algorithm,and to establish the external memory space to store the target area.Moreover,the dynamic clustering strategy is used to cluster the external memory space to determine the target area information,and to complete object detection.The main research content and contributions of this paper include:(1)Propose TDMN(Triple-Deep Measure Net,TDMN)network model.The input of the model is a set of triples composed of positive and negative samples and anchor points,and the distance between positive and negative samples is the output.Design two subnetworks to implement feature extraction and relationship learning of input samples,and provide corresponding learning and training methods.The TDMN network can calculate the metric value between the target example and the input sample,and solve the problem of measurement between the target and the candidate area in the object detection process.(2)Propose a TDMN-PSO object detection method which randomly generates candidate areas in the image to be detected and designs a corresponding coding mechanism to regard each candidate area as an Agent.This method introduces the PSO idea,and use TDMN model to calculate the optimal individual or group value in each iterative search process,which drives the Agent to approach the target area.This method adopts a random search strategy to detect the entire sample of the target detection,which avoids the problem of missed detection and false detection of non-significant targets due to the dependence of existing detection methods on candidate areas.(3)Propose TDMN-AMPSO adaptive object detection,optimizing TDMN-PSO method to have better detection effect in multi-target detection.The traditional PSO algorithm is improved,and the AMPSO algorithm is proposed to generate an external memory space to record the search trajectory of high-quality particles.The AMPSO method can help particles jump out of the local optimal state and expand the particle search field of view.(4)Improve the Agnes clustering method and propose AMAgnes dynamic clustering which does not need to specify the number of cluster clusters.The TDMN metric value is used to constrain the clustering state.The dynamic clustering of the external memory space has been completed and object detection has been achieved after the stable point of the clustering state is determined by evaluation indicators.The method in this paper can be applied to remote sensing images,medical images and other non-significant target data sets with unclear edges.The experimental results have showed that our method can improve the detection recall rate and detection accuracy to varying degrees without relying on candidate areas,which has a strong practical significance.