The accurate quality detection of hardwood logs can realize efficient utilization and profit maximization of wood. However, due to the difference in extraction principle of acoustic parameters and interaction mechanism between parameters and wood properties, the evaluation results in log quality are different to some extent. On this basis, a method for acoustic feature extraction and defect detection is proposed, which is based on the improved particle swarm optimization-variational modal decomposition ( IPSO-VMD). By analyzing the sparse characteristics of defect signals, the minimum average envelope entropy is determined as the fitness function of PSO optimized VMD to search for the optimal combination parameters (K, α). And the search of PSO is accelerated and the global optimal solution is achieved through improving the inertia weight and learning factor of PSO. Then, the effective sub-modes from the IPSO-VMD are selected based on marginal spectrum and energy ratio of sub-band components, and the frequency band distribution and energy ratio of each effective mode are used as the characteristic parameters characterized the defect signal to realize the accurate quality detection of hardwood logs. The actual sawing results show that the major defect types and priorities of logs are detected with accuracy of 88. 6% and 72. 7% , respectively, and which not identifiable by global parameters could be examined effectively based on the IPSO-VMD method. The effectiveness of the new feature parameters can provide a reliable basis for the accurate detection in log quality through fusing the multi-parameter features and constructing the artificial intelligence recognition system.