| Summary: | The development of port automation requires sensors to detect container movement. Vision sensors have recently received considerable attention and are being developed as AI advances, leading to various container motion detection methods. Faster-RCNN is a detection method that performs better precision and recall than other methods. Nonetheless, the detectors are set using the Faster-RCNN default parameters. It is of interest to optimized its parameters for producing more accurate detectors for container detection tasks. Faster RCNN requires mixed integer optimization for its continuous and integer parameters. Efficient Modified Particle Swarm Optimization (EMPSO) offers a method to optimize integer parameter by evolutionary updating the space of each candidate solution but has high possibility stuck in the local minima due to rapid growth of Gbest and Pbest space. This paper proposes two modifications to improve EMPSO that could adapt to the current global solution. Firstly, the non-Gbest and Pbest total position spaces are made adaptive to changes according to the Gbest and Pbest position spaces. Second, a weighted multiobjective optimization for Faster-RCNN is proposed based on minimum loss, average loss, and gradient of loss to give priority scale. The integer EMPSO with adaptive changes to Gbest and Pbest position space is first tested on nine non-linear standard test functions to validate its performance, the results show performance improvement in finding global minimum compared to EMPSO. This tested algorithm is then applied to optimize Faster-RCNN with the weighted cost function, which uses 1300 container images to train the model and then tested on four videos of moving containers at seaports. The results produce better performances regarding the speed and achieving the optimal solution. This technique causes better minimum losses, average losses, intersection over union, confidence score, precision, and accuracy than the results of the default parameters.
|
|---|