Local Estimation In Deep Object Counting

Object counting,an important problem during human life and production,is widely used in security monitoring,intelligent transportation,and agriculture management.It aims to calculate the number of an object.Among the many counting solutions,the computervision-based(CV-based)counting method is a research focus because of its characteristics of high throughput and accuracy.CV-based counting starts from detection or segmentation approaches and is gradually solved by the regression at the global or pixel level.It is still limited by the issues such as data distribution and the noisy ground truth.To solve these issues,the local count approach is proposed.Since the local count framework is robust to the noisy ground truth,it has been one of the mainstream solutions for CV-based counting.However,some problems remain in the local counting,such as sample imbalance,scale variations,and lacking of the instance distinguishability.To address the problem mentioned above,based on the local count framework,this dissertation makes the following contributions:First,considering the issues of the noisy ground truth and sample imbalance in object counting,a strategy based on the dense level classification and a block-wise classification network is proposed.Meanwhile,an information-entropy-based weight is applied to balance the training loss between different dense levels.To tackle the checkerboard artifacts,a redundant post-process is introduced to improve the resolution.The experiments verify that the block-wise classification network significantly outperforms the local count regression baseline.Second,a scale-analysis experiment is presented.It shows a correlation between the object size and the estimation scale,and the performance decreases significantly when they mismatch.Based on the observations above,a scale mismatched suppression mechanism and a scale-mismatch-aware network are proposed to suppress the result predicted by the mismatched scale,only what does not have the obvious estimated error will be fused.Subsequent experimental results show that the proposed method improves the scale adaptation of the local count method.Third,a new solution to model the counting task is proposed.Motivated by the sequential count behavior of humans and the scale weighing process,a counting scale is proposed to ’weight’ the count in an image by a virtual weight.Thus,the counting task is translated into a sequential decision-making task.Further,the counting scale is implemented by different forms of counting scale and sequential decision-making solutions termed LibraNet series networks.Experiments built on several object counting benchmarks analyze the characteristics of LibraNet series networks,and the best or at least competitive performance of the LibraNet series is reported compared with the mainstream counting solutions such as the density-map-based methods.Based on the research of common object counting,local counting is implemented into real scene applications.To fill the gap between the local count framework and the distinguishability of instance required by agricultural object counting,the last work of this dissertation focuses on improving the distinguishability of instance of the local count framework.An analysis reveals that the local counting lacks the distinguishability of instance because of low-resolution output and the error response in the background.To solve the problems above,a guiding upsampling process and a counting strategy based on background prior are proposed,which are further fused into the local count framework to present an instancedistinguishable plant counting method.The experiments built upon several plant counting benchmarks show that the output of the proposed method shows the instance more distinguishably,and the proposed method significantly improves the counting performance.These experimental results verify that the proposed method is an instance-distinguishable and accurate plant counting approach.